Date: 2026-02-27 Author: Research and proposal for Mariana's review Last news post: January 18, 2026 (Cinevva launches reels-first game discovery) Gap: 40 days without a news update

Part 1: Catch-up news articles

These are factual news posts following the existing format. Short, direct, developer-focused.

1. Godot 4.6 released with Jolt Physics and new editor theme

Why it matters: We already published a preview article on January 10 predicting this release. Now it's shipped and the actual features are bigger than expected. Jolt Physics is now the default for all new 3D projects, replacing the old GodotPhysics. Direct3D 12 is the default renderer on Windows. There's a new IK framework, SSR overhaul, and a fresh "Modern" editor theme. LibGodot lets you embed the engine in other applications. Delta encoding makes patch updates smaller. This is one of the most feature-dense Godot releases in years and it directly affects web game developers.

Angle: Follow up on our preview. What shipped vs. what we expected. Emphasis on Jolt Physics (matters for web export performance) and LibGodot (embedding potential).

Date to publish: 2026-02-27

2. Unity previews AI game creation ahead of GDC

Why it matters: Unity CEO Matthew Bromberg announced that Unity will demo a new AI beta at GDC (March 9-13) that lets people "prompt full casual games into existence with natural language only." This is a direct competitor to what Cinevva already does. It's also a signal that the biggest engine in the industry now considers AI-first game creation a primary product direction, not a side experiment.

Angle: Factual reporting on the announcement. What Unity said. What we know and don't know. The significance of a mainstream engine going all-in on prompt-to-game. No editorializing in the news post itself.

Date to publish: 2026-02-28

3. Roblox launches AI-powered 4D model creation

Why it matters: Roblox released a beta tool that generates fully functioning in-game 3D models from natural language prompts, complete with physics and interactive mechanics. Cars with working doors. Objects with real collision. This is relevant to our audience because it shows AI-generated game content moving from static assets to interactive objects.

Angle: What the 4D creation tool does. How it compares to existing 3D generation tools. What "4D" means in this context (3D + behavior/interactivity).

Date to publish: 2026-03-01

4. Godot maintainers struggle with AI-generated pull requests

Why it matters: Rémi Verschelde, one of Godot's core maintainers, called the flood of AI-generated PRs "draining and demoralizing." Contributors are submitting low-quality, nonsensical code generated by AI tools without review. This is affecting one of the most important open-source game engines and it connects to a broader issue that matters to our community.

Angle: The facts. What's happening to Godot's PR queue. Verschelde's direct quotes. The tension between welcoming new contributors and managing noise. Keep it factual for news, but this topic could also warrant a deeper piece in the new section (see below).

Date to publish: 2026-03-03

5. Steam Next Fest February 2026 data: what worked for indie discovery

Why it matters: Over 2,200 games participated. The data from presskit.gg shows clear patterns: games with 10,000+ existing wishlists broke into the Diamond tier. Most games earned 460-999 wishlists. Co-op and survival-craft overperformed. Narrative games and visual novels struggled. This is directly relevant to our audience of indie developers trying to get discovered.

Angle: The numbers. What tiers existed. What genres performed well. What the data tells you about when to enter Next Fest and how to prepare. Reference our existing Steam Next Fest strategy guide.

Date to publish: 2026-03-05

Part 2: New section proposal — "Signals"

The concept

A new content section called Signals that sits between News and Blog.

News is factual. What happened. Short. No opinion. Blog is personal. Deep essays. Mariana and Oleg writing from lived experience. Signals is analytical. What does this trend mean? Where is this going? Written in Mariana's conversational voice but focused on industry shifts rather than personal stories.

Why it fills a gap

The news section reports events. The blog section tells stories. Neither one does the thing readers increasingly want: someone they trust telling them what to pay attention to and why. That's what Signals does.

Each piece is 500-900 words. More frequent than blog posts (which take weeks). More opinionated than news (which stays neutral). Published when something worth analyzing happens, not on a schedule.

Voice and format

Written in Mariana's voice following the existing guidelines:

• Conversational, direct, warm
• First person ("We've been watching..." / "I think...")
• Contractions, active voice
• No corporate speak, no banned words
• Specific and concrete, not abstract
• Opens with a hook, not a summary
• No bullet points unless the structure demands it
• No summary paragraph at the end

Each Signals piece follows a loose structure:

1. The thing that happened (1-2 sentences)
2. Why it caught our attention (the analysis)
3. What it means for people building games (the "so what")

No formal sections or headers required. These should read like smart notes from someone paying close attention.

Proposed Signals articles

Signal 1: "Everyone wants to be the AI game engine now"

Unity is demoing prompt-to-game at GDC. Roblox shipped 4D model creation. Moonlake AI opened its World Model Agent beta. Phaser Editor v5 integrated AI through MCP. Arcane Forge is building a full AI pipeline. And we've been doing this for years.

This isn't a complaint about competition. It's an observation about timing. Two years ago, VCs said the market wasn't there. Now the biggest names in gaming are racing to build what they said couldn't work.

The piece would cover what each company announced, how the approaches differ (text-to-game vs. AI copilot vs. AI asset pipeline), and why the "AI game creation" umbrella hides fundamentally different products. Some of these tools generate a demo that looks cool in a tweet. Others generate something you can actually ship. The difference matters.

Target length: 800 words

Signal 2: "Vibe coding is the new game jam"

Collins Dictionary named "vibe coding" Word of the Year 2025. Andrej Karpathy coined it to describe building software by describing intent in plain English. In game development specifically, this means the barrier between "weird idea" and "playable prototype" collapsed.

This piece would connect vibe coding to something we already demonstrated with A Breaker Belt (built in three days). The point isn't that AI writes better code than humans. The point is that game ideas that would have died in a brainstorm doc now get built in an afternoon. The economics of experimentation changed.

The risk of vibe coding: context window limits, loss of understanding, AI-generated complexity that nobody can debug. The promise: more games that shouldn't exist but somehow do.

Target length: 700 words

Signal 3: "Open source has an AI pollution problem"

Godot maintainers are drowning in AI-generated PRs. The code doesn't work, doesn't make sense, and wastes volunteer time. Rémi Verschelde called it "draining and demoralizing." This is happening to one of the most important open-source game engines in the world.

The piece would explore the irony: AI tools that are supposed to make developers more productive are actively making open-source projects less productive. The people submitting these PRs aren't malicious. They're using the tools they've been told to use. The incentive structure is broken. AI makes it trivially easy to generate a PR. It doesn't make it easy to generate a good one.

This connects to the broader question of AI in creative work. Making something is now cheap. Making something good still requires taste, judgment, and effort. The tools changed. The standard didn't.

Target length: 700 words

Signal 4: "What 2,200 games at Steam Next Fest tell you about discovery"

Steam Next Fest February 2026 is a dataset, not just an event. 2,200+ games. Clear tier breakdowns. Genre performance data. Patterns you can actually act on.

This piece would go deeper than the news article (which covers facts). It would analyze what the data means for indie developers choosing where and how to launch. Co-op games overperformed. Narrative games underperformed. Games with existing momentum compounded it. Games starting cold mostly stayed cold.

Then connect it to what Cinevva is building. Steam Next Fest rewards games that already have an audience. Cinevva's reels-first discovery is designed for games that don't. The approaches are complementary, not competing. But the Next Fest data shows why alternative discovery paths matter.

Target length: 800 words

Section logistics

URL structure: /signals/ with individual posts at /signals/YYYY-MM-DD-slug Frontmatter: Same as news (title, description, date) plus author field like blog Index page: Similar to news index, chronological, with RSS Navigation: Add under the "Learn" dropdown alongside News, Blog, Tutorials, Guides Frequency: 2-4 per month, whenever something worth analyzing happens Authors: Primarily Mariana. Oleg for technical signals. Guest contributors eventually.

Part 3: About "Alina"

I searched the entire codebase, blog posts, about page, team pages, and external sources for references to "Alina." I found no one by that name associated with Cinevva or the immediate game development tool space.

Possibilities:

• Alina.social is a social media content management platform with an AI editor, but it's not related to game development
• There's an "Alina" user on GDevelop (gd.games/Alina) who creates games on the platform
• Could be a team member, collaborator, or friend whose work isn't yet reflected in the codebase

If "Alina" refers to someone specific on the team or in the community, I'd need more context to research what she's building and incorporate it into the content plan.

Part 4: Recommended publishing order

Week of Feb 27 - Mar 2:

1. News: Godot 4.6 released (Feb 27)
2. News: Unity previews AI game creation (Feb 28)
3. News: Roblox launches 4D AI models (Mar 1)

Week of Mar 3 - 9: 4. News: Godot maintainers and AI PRs (Mar 3) 5. Signals: "Everyone wants to be the AI game engine now" (Mar 4) 6. Signals: "Open source has an AI pollution problem" (Mar 6)

Week of Mar 10 - 16 (GDC week): 7. News: Steam Next Fest February results (Mar 10) 8. Signals: "What 2,200 games tell you about discovery" (Mar 11) 9. Signals: "Vibe coding is the new game jam" (Mar 13)

This schedule gets us back to regular publishing, establishes the Signals section, and creates a rhythm around GDC week.

Moonshot AI launched Kimi K3 on July 16 and released the weights eleven days later, on July 27. The download is 96 shards and roughly 1.56 TB on Hugging Face. Nathan Lambert called it "clearly the strongest open model ever released," and the rankings back him up: first place in the Frontend Code Arena, second on the Vals AI Index, third on the Artificial Analysis Intelligence Index behind only Claude Fable 5 and GPT-5.6 Sol Max. Among open-weight models it isn't close.

That's the story everyone wrote. The part that caught our attention is what Moonshot chose to show off with it.

The demo is a Three.js game

Alongside the launch, K3 produced a procedural 3D exploration game that runs in a browser tab: an open world with water, trees, cabins, snowy mountains, and riders on horseback, built in Three.js on the WebGPU renderer with GPU compute. Not a research clip, not a video of a game. A page you load.

The method matters more than the scene. According to the WebGPU community's writeup, the model worked with vision in the loop, iterating between the code it generated and live screenshots of the running result so it could see what it had made and fix it. That is the same loop we run inside Cinevva's Game Creator, and it is the thing that separates a model that writes plausible Three.js from a model that ships a game that actually renders. Anyone can generate a scene graph. Noticing the camera is inside the terrain is the hard part.

So the situation as of this week is that a frontier lab, launching its most capable model, reached for browser-based 3D game creation as the demonstration that would land. Two years ago the flagship demo was a chatbot answering a riddle. The bar moved to something you can walk around in.

What's actually in the model

K3 is a Mixture-of-Experts with 2.8 trillion total parameters and 104 billion active per token, a context window of 1,048,576 tokens, and text, image, and video input. Moonshot introduced two architecture changes, Kimi Delta Attention and Attention Residuals, and reports roughly a 2.5x improvement in scaling efficiency over Kimi K2. It always reasons, and where the July launch shipped with a single "max" effort level, the open release exposes low, high, and max.

On Moonshot's own numbers it beats Claude Opus 4.8 at max and GPT-5.5 at high across most of the suite, and loses to Fable 5 and GPT-5.6 Sol. Treat self-reported benchmarks as a claim rather than a verdict, but two results held up in independent testing and both are the kind that matter for building software: BrowseComp at 91.2 for agentic browsing, and SWE Marathon at 42.0 for long-horizon coding, which puts it above both Fable 5 and Sol on work that runs for hours instead of seconds.

API pricing is $3.00 per million input tokens and $15.00 output, dropping to $0.30 on cached input, the same headline rate as Claude Sonnet 5. OpenRouter is listing it slightly cheaper. One caveat worth reading before you build on it: unlike Tencent's Hunyuan Hy3, which went out under Apache 2.0 with no restrictions, K3 ships under a custom Kimi K3 License. Open weights and open source are not the same document, and if you plan to put this in a commercial product you should read the terms rather than assume.

The gap is closing faster than the release cadence suggests

Lambert's read on the broader trend is that the distance between the best closed models and the best open ones has compressed from six to nine months down to something more like three to five. Alibaba has already announced Qwen 3.8 at 2.4 trillion parameters with open weights to follow, so the escalation isn't slowing. We wrote in January 2025 that DeepSeek R1 reset the cost of intelligence. Eighteen months on, the pattern is that every reset gets absorbed and then repeated faster.

For anyone making games, the practical consequence is that the intelligence layer keeps getting cheaper and more portable at the same time the asset layer already did. A studio that wanted a model fine-tuned on its own codebase used to be choosing between an API dependency it didn't control and a weaker local model. That tradeoff is thinner every quarter.

What we're doing about it

We run a multi-provider gateway, so this is a concrete decision rather than an abstract one. K3's strengths sit exactly where our workload is: long-horizon agentic coding, tool use, and iterating against screenshots and console logs from a live game. We're benchmarking it against our current model lineup on real Game Creator sessions rather than on anyone's published table, because the metric we care about isn't an Elo score. It's how often a turn ends with a game the person can actually play, and how often the model catches its own mistake before the user has to.

We'll publish what we find, including if the answer is that it doesn't beat what we're already running. The interesting thing about this release isn't that a Chinese lab shipped a very good open model, which has stopped being surprising. It's that the demo they picked to prove it was a game in a browser tab, which is the thing we've been building for.

References

• Simon Willison: Kimi K3, and what we can still learn from the pelican benchmark
• Nathan Lambert: Kimi K3, the open-weights escalation
• WebGPU community: Kimi K3 uses Three.js WebGPU for an AI-built 3D game demo
• OpenRouter: moonshotai/kimi-k3
• explainX: Kimi K3 open weights, 2.8T parameters, day-0 hosting

Black Forest Labs released FLUX 3 on July 23, and it's a bigger jump than the version number suggests. The Freiburg lab built its name on FLUX, the image models that generate art inside Photoshop, Canva, Picsart, and, closer to home, inside Cinevva for everyone who picks the Flux provider. FLUX 3 is the first time the company ships video at all, and it doesn't ship a separate video model. It trains one set of weights on images, video, and audio at once, then extends that same backbone to predict robot actions. One model, four kinds of output.

A first look at what FLUX 3 Video generates

One backbone, not four products glued together

The pitch rests on a research idea BFL calls Self-Flow: instead of learning image generation, video generation, and audio generation as separate problems, learn a single representation of how the world works. An image is a slice of space at one moment. Video adds time and motion. Audio carries the causal link between an impact and the sound it makes. Train on all of them together and the constraints reinforce each other, so the sound has to match the motion and the motion has to obey physics. The surprise in the launch is that action prediction fell out of the same backbone without wrecking the video quality, which is why FLUX 3 can drive a robot arm and generate a music video from the same weights.

The video, concretely

FLUX 3 Video makes clips up to 20 seconds long in a single pass, with native audio generated alongside the picture, dialogue, sound effects, and ambient noise included. That length matches OpenAI's now-shut-down Sora and beats most of the current field on duration. It handles text-to-video, image-to-video, video-to-video, keyframe transitions, multilingual dialogue, and agentic chaining of clips into multi-shot sequences. In BFL's own early tests on 10-second 720p clips, reviewers preferred it over Luma Ray 3.2 in 93% of comparisons and Runway Gen-4.5 in 77%. Against the strongest rivals the margin shrinks to a coin flip: 52% over both ByteDance's Seedance 2.0 and Google's Gemini Omni Flash. These are the lab's own preliminary numbers, so treat them as a claim, not a verdict.

The catch is access. FLUX 3 Video is in gated early access through the API and private weights, and you have to apply and get approved. FLUX 3 Action is partner-only, already being tested on a real Audi production line through robotics startup mimic. FLUX 3 Image, the tier most designers actually want, lands in the coming weeks. The open-weight FLUX 3 Dev backbone, the piece the local-inference crowd cares about, isn't due until later in 2026.

Why an image lab going multimodal matters to us

We watch BFL closely because Cinevva already runs on their work. When one of the two labs whose image models sit inside the world's biggest creative tools decides its future is a unified video-audio-action model, that's a signal about where the whole creation stack is heading, not just a new toy. The line between "generate a picture," "generate a clip," and "simulate a world you can act in" is dissolving into one model, and it's the same convergence we flagged with the MIRA world model from the other direction.

For creators the practical read is steady. A better, longer, audio-native video model lowers the cost of the first draft again, same as Seedance, Kling, and Gemini did before it. What it doesn't touch is the part that was always hard: knowing what's worth making, then iterating a generated clip into something people actually want to watch or play. That's the loop we keep building around, from the prompt to sharing, remixing, and publishing on the open web. FLUX 3 makes the raw material cheaper and better. Turning that material into a game or a story people care about is still the work.

References

• Black Forest Labs: FLUX 3, Real World Models
• VentureBeat: Black Forest Labs launches FLUX 3
• Decrypt: FLUX 3 ditches stills for video and robot hands
• The Decoder: Flux 3 generates videos with native audio up to 20 seconds long
• Black Forest Labs press release (GlobeNewswire)

Roblox announced Build on July 16. Type "let's make a cozy adventure game set in a dense forest with environmental obstacles" into the Roblox mobile app and it generates a starting point you can playtest, iterate on, and publish. Public alpha opens July 28 in New Zealand for age-checked users nine and older. This is the biggest user-generated-content platform in the world deciding that a sentence typed on a phone is now a first-class way to create a game.

What Roblox actually announced

Build lives inside the Roblox app itself, not Studio. The prompt drives gameplay mechanics, environment, characters, visual style, and sound, powered by a mix of open-source models and proprietary Roblox models trained on 3D and gaming-specific data. There's a free base version with paid options for heavier use. Creations go through the standard safety review, and publishing to the global catalog requires an age-checked account that's 16 or older, a notably harder gate than the nine-plus bar for just making things.

Build isn't shipping alone. Roblox also previewed a playtesting agent that hunts bugs before launch, an analytics agent you query in plain language, an experiment agent that proposes engagement tests, parametric 3D assets from text or images, and a scene-generation model that assembles complete 3D environments. Taken together it's less a feature and more a pipeline: generation, testing, tuning, and analytics, all agentic, all inside one platform.

Why the mobile part matters

Roblox's audience is overwhelmingly on phones, but its creation tools never were. Studio requires a desktop, and that requirement has quietly filtered who gets to be a creator for close to two decades. Build removes the filter. Most of Roblox's youngest and most engaged users are about to get creation tools in the same app they already play in, with no laptop in the household required.

That's the same bet we made with Cinevva running entirely in the browser, and it's the right one. The interesting wrinkle is the age split. Nine-year-olds can build but only 16-plus users can publish globally, which means Roblox expects an enormous volume of creation that never leaves the sandbox. That's a training-data and habit-formation play as much as a content play.

The walled garden question

Everything Build makes is a Roblox experience. It runs on Roblox, monetizes in Robux under Roblox's revenue share, and lives or dies by Roblox's retention-based discovery. The prompt is yours, the game isn't portable. That's not a criticism so much as the whole strategy: Roblox has said it wants 10% of all gaming content revenue flowing through its platform, and an AI creation layer that multiplies content supply is how a marketplace gets there.

For creators the trade is real. You get Roblox's 380 million monthly users as a potential audience and give up ownership of the artifact. A web game you can host anywhere, embed anywhere, and keep when the platform changes its terms is a different kind of asset than an experience inside someone else's catalog.

What it means for Cinevva

Build is the loudest validation yet that prompt-to-game is the next creation interface, and it will teach millions of kids the grammar of describing a game in words. We think that's great for everyone in this space. Our answer stays the same as it was last week: games you make with Cinevva are real web games, yours to publish on the open web, port to WeChat, or ship anywhere a browser runs. Roblox built an AI factory inside its walls. We'd rather hand you the factory.

References

• Roblox: Build Without Limits on Roblox
• TechCrunch: Roblox launches an AI-powered game-creation feature in its mobile app
• PocketGamer.biz: Roblox unveils mobile-first AI game creation with new Build tool
• The Next Web: Roblox launches Build, a mobile tab that turns text prompts into playable games
• Engadget: Roblox will offer AI-generated game creation on mobile

Sulka Haro, the former Habbo Hotel lead, published a full census of 53,597 Steam releases covering mid-2023 to mid-2026. The headline number: roughly one in three new Steam games now carries Valve's AI-content disclosure, up from about 7% when the mandatory field launched in early 2024. AI-flagged launches grew from 13 a month to around 530, and they account for 60 to 90 percent of all growth in Steam's monthly release count. If the trajectory holds, AI-disclosed games cross half of all releases sometime in 2027 or 2028.

The supply explosion is real

The year-over-year curve is steep and clean. AI-flagged games were 10.9% of releases in 2024, 19.9% in 2025, and 30.8% so far in 2026. Strip them out and Steam's much-discussed release growth mostly disappears. The tooling that collapsed asset costs and code costs didn't just make existing developers faster, it minted new ones, and as Haro puts it, AI lowers the barrier not just to make a game but to make several.

One caveat he's careful about and most coverage isn't: the disclosure is binary. A game that ran one texture through an upscaler and a game that generated its entire art set carry the same flag. It's a touched-AI signal, not a dose measurement.

The revenue gap

Here's the uncomfortable half of the study. AI-disclosed games are a third of new releases but only somewhere between 10 and 27 percent of estimated sales in the most recent stretch, up from 3 to 6 percent in 2024. Growing, but lagging supply badly. And Steam's brutal baseline still applies to everyone: the top 1% of titles capture about 94% of estimated revenue. AI made shipping a game dramatically cheaper. It did nothing to make anyone want to play it.

What the successful AI games do differently

The most useful finding is buried in the usage breakdown. Among AI-flagged games that failed to find an audience, 72% used AI for visuals, the most visible and most resented application. The successful ones lean on AI where players don't stare at it: voice at 24% versus 8%, localization at 18% versus 6%, plus text and music. Their disclosures also read differently, full of words like "auxiliary" and "reviewed and refined." The winners use AI as a multiplier on craft. The flood uses it as a substitute.

What it means for creators

This study is the clearest data yet on the thing we see every day: creation is no longer the bottleneck, attention is. Half a thousand AI-assisted games a month are landing on a store where 1% of titles take 94% of the money. The lesson isn't to avoid AI, the successful cohort uses it more broadly than the failures. The lesson is that generation gets you to the starting line, and iteration, playtesting, and distribution decide the race. That's why Cinevva keeps investing in the loop after the first prompt, sharing, remixing, and publishing to the open web, and not just the generation moment. Making a game is now easy. Making one that people play is the entire game.

References

• Sulka Haro: Three years of AI on Steam
• GamesRadar: Half of all games released on Steam will use AI by 2028, study predicts
• GamesRadar: Steam study of over 53,000 games finds 60 to 90 percent of release growth comes from AI games
• PC Guide: More than 17,000 games on Steam now come with an AI disclosure

WeChat mini games passed 500 million monthly active players in 2025, with over a billion registered users and roughly half a million developers building for the platform. The Chinese mini game market did 39.8 billion yuan (about $5.6 billion) in 2024, nearly doubling year over year. It's the largest instant-games audience in the world, and almost none of it is reachable by the web games most of you build with us.

So we're building a path to it. Starting this week, Cinevva's Game Creator is rolling out an experimental WeChat mini game mode.

What the mode does

WeChat mini games don't run in a browser. They run in a stripped JavaScript environment inside the WeChat app: there's a canvas and a set of wx.* APIs, and that's it. No DOM, no CSS, no HTML buttons, no Pointer Lock, no service workers. Packages are capped at 4 MB for the main bundle and 20 MB total with subpackages. And the platform bans dynamic code outright, so a game can't download or eval a single line of JavaScript at runtime.

That last rule is the one that shapes everything. It means there's no such thing as a "Cinevva player" inside WeChat that streams your game from our servers. Every game has to be compiled into its own standalone package and submitted for review on its own.

The new mode prepares your game for exactly that from the first prompt. When it's on, the Game Creator builds within the portable subset: all UI is drawn on canvas instead of HTML, controls are touch-first with drag-look instead of pointer lock, rendering is budgeted for a mid-range phone with a boot gate for devices that lack real WebGL2, assets stay small and audio ships as mp3. The game still runs in your browser like any other Cinevva game. It just doesn't pick up any of the habits that would make a WeChat port a rewrite.

The exporter that turns one of these games into a ready-to-submit WeChat DevTools project is the next step, and it's in progress. Mode first, packaging second, because the mode is what determines whether packaging is mechanical or hopeless.

The rules, honestly

Building a portable game is the easy half. Publishing in China is regulated, and you should know what you're walking into before you invest.

Every mini game needs an ICP filing, which has been mandatory since late 2023 and takes a week to a few weeks. Every title also needs proof of copyright, either a software copyright registration or the faster electronic copyright certification, which takes roughly two weeks per game.

Monetization is where the hard line sits. Any form of in-app purchase requires a game publication license, the banhao, and only a Chinese company can apply for one. If you're a developer outside China, that means partnering with a licensed domestic publisher, and imported titles go through a slower approval track that has historically taken six months or more.

There's one path that doesn't require a banhao: free games monetized through WeChat's own ad components. That's not a loophole, it's how a large share of the mini game market actually operates. Ads were 31% of mini game revenue in 2024, and for a developer outside China it's the realistic way to test the market before committing to a publisher relationship.

Where this is going

The mode is experimental and we're being upfront about that. We're using it ourselves first, tightening the generation rules against real device testing, and building the export pipeline behind it. If you're interested in reaching Chinese players with a Cinevva game, or you're already navigating a mini game launch and want to compare notes, tell us. The fastest way to shape what this becomes is to be one of the first people using it.

References

• PocketGamer.biz: WeChat Mini Games crosses 500m monthly active users
• iChongqing: WeChat Mini Games hits 1B users, becomes global magnet for game developers
• Oreate: 2025 WeChat mini game ecosystem development report
• WeChat official docs: JavaScript support in mini games
• Niko Partners: game regulations in China
• AppInChina: the complete guide to WeChat mini games

Earlier this month, the Godot Foundation published an update to its contribution policies, dated June 30, that draws the hardest line any major engine has drawn on AI. All code contributed to the engine must be human authored. AI assistance is allowed only for menial work like code completion, regex, or find-and-replace, and if you used AI in any capacity to author code, you have to say so in the PR discussion. The same rule covers communication: issue descriptions, PR messages, and proposals must be written by a person, not generated by a model. Autonomous agents and vibe-coded submissions were already grounds for an automatic ban from the repository, and they stay that way.

Community coverage of the Godot contribution policy update

The policy, in plain terms

The reasoning the Foundation gives is about responsibility, not ideology. A contribution to Godot isn't a one-time transaction. Whoever writes the code needs to understand it well enough to defend it in review, fix it when it breaks, and answer questions about it years later. The post puts it bluntly:

AI can't take responsibility for a line of code, so a human has to, and a human who didn't write the code usually can't. The communication rule follows the same logic. Maintainers volunteer their time to review your work, and the Foundation calls it a basic principle of respect that they get to talk to a person rather than a machine. Machine translation still passes, as long as a human wrote the original. There's also a new brake on ambition: contributors with three or fewer merged PRs can't open new features or significant refactors without a maintainer's go-ahead.

Four months from complaint to policy

None of this arrived out of nowhere. Back in March we covered Rémi Verschelde describing the flood of AI slop PRs as draining and demoralizing for maintainers, with the open PR queue sitting at 4,681. By the time the policy landed, coverage put the queue above 5,000. The economics haven't changed since March. AI dropped the cost of generating a plausible-looking PR to near zero while the cost of reviewing one stayed exactly where it was. What changed is that Godot stopped treating that as a moderation problem and started treating it as a policy problem. The update came less than two weeks after Godot 4.7 shipped, which tells you the maintainers' scarce resource isn't the ability to ship, it's review attention.

Contribution isn't creation

We build an AI game-creation platform, so let's be direct about the tension instead of pretending it isn't there. The most-used open-source indie engine just said AI-written code isn't welcome, and AI is the whole way games get made on Cinevva. Those two things sound opposed. They aren't, and the difference is who carries the code afterward.

When you merge a PR into Godot, hundreds of thousands of developers inherit that code, and a small group of volunteers becomes responsible for it indefinitely. The author's understanding is the only warranty the project gets, which is exactly what heavy AI use removes. When you build your own game with AI, you own the result. If it breaks, it's your game that breaks. Nobody else inherits your code, and no volunteer has to review it.

Godot didn't ban making games with AI, and couldn't if it wanted to. It banned outsourcing your half of a shared responsibility to a model. That's a defensible line, and honestly, it's the same one we'd draw. The tools can write the code. Someone still has to stand behind it.

References

• Godot Engine: Changes to our Contribution Policies
• PC Gamer: Open source game engine Godot will no longer accept AI-authored code contributions
• Game Developer: Godot to ban (almost all) AI coding contributions
• Kotaku: Indie game engine will require all code to be human-authored

On July 3, Business Standard published a report on how generative AI is rewriting the economics of making games in India, and it's worth reading past the headline. The claim isn't that AI writes good games. It's that AI is collapsing pre-production costs inside Indian studios, and pre-production cost is exactly the thing that has kept India a country of players rather than a country of game makers. Parth Chadha, CEO of the gaming platform STAN, gave the report its sharpest line: "Building globally scalable IP takes 50 swings to land one hit, and the cost of each swing was historically too high."

BGMI's official launch video from 2021. The game turned five on July 2, 2026, with 260 million downloads.

The 50-swings problem

Fifty swings per hit isn't a complaint, it's a budget line. If each attempt at a new IP costs a studio months of concept art, prototype code, and design iteration, a market without deep capital never gets enough at-bats to produce a hit. Chadha's read is that AI compresses the cost of every one of those swings, so the bottleneck moves from capital to taste. Anuj Tandon, who leads Bitkraft Ventures' investments in India, makes the parallel argument in the same report: AI is dropping the barrier across art, code, localization, and design all at once, which means a small Indian team can now attempt what used to take big-studio headcount.

This isn't theoretical. Open magazine's coverage of the ecosystem shows where it's already landing. Hyderabad's Yesgnome demoed Sketly AI, which turns character, environment, and animation asset work from weeks into minutes. Metasports runs AI-powered live commentary inside Hitwicket, its cricket strategy game with over 18 million players. The pre-production line items that used to be the expensive part are becoming the cheap part.

A huge audience, an industry pushed to pivot

The context makes the timing sharper. BGMI, India's biggest game, turned five on July 2 with 260 million downloads. That number is the whole paradox in one figure. India consumes games at global scale, but its biggest title belongs to Krafton, a Korean company. The player base was never the problem. The economics of building for it were.

And the money side of the industry just got shoved toward content whether it wanted to or not. The 2025 ban on real-money gaming ended the fantasy-sports engine that absorbed most of Indian gaming's capital and talent for a decade. Dream11's fintech arm Dream Money winds down on July 30, less than a year after launch, one of the clearest signs that the RMG era's diversification plans are being rolled back rather than rescued. What remains for the ecosystem is the thing it mostly skipped on the way up: making actual games. The falling cost of a swing arrives at the exact moment India needs a lot of swings.

The cheap swing

We've been tracking this shift from the solo-dev side for a while, and the pattern is the same everywhere: when production cost collapses, the number of attempts explodes, and hits come from wherever attempts are cheapest. That's the entire idea behind Cinevva. A browser, a prompt, a playable game you can share in an afternoon is about the cheapest swing that exists. If Chadha's math is right and IP is a numbers game, then a country with 260 million downloads of proven demand and a newly affordable at-bat is exactly where you'd expect the next wave of homegrown games to come from.

References

• Business Standard: The cost of creation, AI is rewriting the economics of making games
• Open: How AI is transforming India's gaming and creative industries
• InsideSport: BGMI 5th anniversary, date, event, celebration and more
• Business Standard: Dream Money shuts after RMG ban ends Dream11 fintech foray

ICML, one of the world's top machine learning conferences, ran at COEX in Seoul from July 6 to 11, drawing around 15,000 attendees and a record 23,918 paper submissions. Korea's game industry treated it like a home game. On opening night, Krafton and the world-model startup Odyssey co-hosted the official AI for Games event, and about 500 researchers and industry people showed up. Krafton also landed 10 main-track papers, 20 counting workshops, its biggest haul ever at a top AI conference. Then, two days after ICML closed, four smaller Korean firms ran a free conference teaching indie devs how to ship games with AI. One city, one week, and the whole industry from research lab to solo dev moved in the same direction.

PUBG Ally, Krafton's co-playable AI teammate, revealed with NVIDIA at CES 2025 and now live in PUBG as ALLIE

Three fronts, one aimed at engines

Krafton's chief AI officer Lee Kang-wook laid out three fronts. The first is in-game AI agents, and PUBG ALLIE is the flagship: an AI teammate you talk to and squad up with, with latency low enough that, in Lee's words, it feels almost like a real friend. A three-layer memory stack, in-context memory, in-session summaries, and cross-session retrieval, lets it remember you across matches. Krafton trained it on data gathered from thousands of internet cafe players to make it behave more like a human. The second front is interactive world models, which Lee positioned as a possible challenge to traditional game engines. The third is production AI for the pipeline itself, and Lee was blunt about why: a single game today takes four to five years and massive costs.

The panel pushed back on the engine-versus-world-model framing in a useful way. A game engine is already a transition function, so in a sense everything is a world model, and the likelier future is neural networks integrated with physics and graphics engines rather than replacing them. Kim Min-jae, CTO of NCSOFT's AI arm NC AI, added the counterweight of the night: "AI is not the protagonist of the stage, but a supporting actor." The protagonist is always the gameplay.

Two days later, the indies got their turn

AMPLIFY 2026 ran on July 13 at D.CAMP Seolleung in Seoul, co-hosted by Com2uS Platform, Anchor Node, UModeler, and Changjo Gongjakso, free for the first 130 registrants, with Unity Korea and the Korea Mobile Game Association sponsoring. No keynote theatrics, just working sessions: how Anchor Node shipped nine games in two years with AI tools, AI 3D asset workflows, AI sound design, and taking a game global. The tagline was "overcoming my limits with AI."

The context makes it land harder. KOCCA data shows generative AI adoption among Korean game companies hit 70% in the fourth quarter of 2025, up from 41.7% just two quarters earlier, and the government is now subsidizing AI tool subscriptions for about 500 small studios. This is also the country that already requires AI-content labels in games by law. Korea isn't debating whether AI belongs in game development anymore. It's teaching the workflows in a free event hall above a startup hub.

What it means for builders

The interesting tension in Seoul wasn't AI versus no AI. It was Krafton's research Seoul, where world models get framed as a rival to game engines, against AMPLIFY's indie Seoul, where AI is just the way a two-person team ships nine games. Both point at the same collapse: the four-to-five-year, massive-cost pipeline Lee described is exactly what's dissolving. That's the ground Cinevva builds on. We think the world-model thesis is right in spirit, the engine stops being the moat and generation becomes the workflow, and we're pushing that all the way down to a browser and a prompt. And Kim Min-jae's line is the right compass for everyone building here: the AI is the supporting actor. If the gameplay isn't the protagonist, no amount of infrastructure saves you.

References

• Inven Global: AI for Games, KRAFTON highlights 'three fronts'
• Seoul Economic Daily: Krafton lands 10 papers at top AI conference ICML
• Inven Global: Four domestic game-AI firms to co-host AMPLIFY 2026 conference
• Khan Games: Four Korean game-AI companies host AMPLIFY 2026
• Seoul Economic Daily: Public, private sectors launch AI aid, investment fund for game firms

General Intuition, Kyutai, and Epic Games released MIRA in early July, and it's the kind of thing that sounds fake until you play it. MIRA is a 5-billion-parameter world model that simulates a full 2v2 Rocket League match with no game engine underneath. No physics code computes the ball's bounce, no renderer draws the arena. A latent diffusion model takes the button presses of all four players and generates the next frame of every player's view, 20 times a second, on a single GPU. The code is on GitHub, the training data is on Hugging Face, and there's a live demo you can play in your browser at mira-wm.com right now.

The multiplayer part is the news

Playable neural simulations aren't new. GameNGen ran Doom inside a diffusion model back in 2024 and Oasis did the same for Minecraft. But those were single-player. The model answered to one person and treated everything else as scenery. MIRA answers to four people at once. When one car touches the ball, all four screens have to agree on where it goes, and the model has to work out whose input caused which change. It generates four synchronized first-person views of the same match and keeps them coherent enough that an actual 2v2 game works. That's the hard problem here, and as far as we can tell, nobody has shipped an open version of it before.

The training data is a release of its own. The Rocket Science dataset holds roughly 10,000 hours of 2v2 matches, about 29 TB across more than 15,000 games, with time-aligned 720p video, keyboard inputs, game events, and physics state logged at 120 Hz from all four players' perspectives. Every match was played by bots, specifically the publicly available Nexto bot, so there's no human gameplay in it at all. The code ships under Apache 2.0, the dataset under CC BY-NC-SA 4.0, with commercial use of Rocket League content still needing Epic's separate permission. And Epic's name on this is the detail worth pausing on. The company that sells the world's most famous game engine just helped publish a playable game with no engine in it.

What it can't do yet

The authors are refreshingly upfront about the limits. That "single GPU" is an Nvidia B200, a datacenter card, not something in anyone's gaming rig. The output is 720p at 20fps, well below what the real game does on a mid-range PC. The context window is short, so the model can lose track of the score and the clock. The ball sometimes drifts toward a goal on its own, cars occasionally twitch without input, and rare events are undertrained because bots rarely produced them. It simulates one game, learned from one bot's style of play. None of that diminishes the result. It just tells you where on the curve we are.

What it means for game creators

Look at the slope. In 2024 a world model ran Doom for one player. In 2025 it was Minecraft. In mid-2026 it's a four-player team sport with vehicle physics, released as open code with data and a browser demo. Engines aren't in danger this quarter, since a B200 per match against a browser tab is no contest. But two lines are converging: engines are absorbing AI-generated content at runtime, like Meshy's runtime-generated mechanics, while neural simulations are becoming playable, multiplayer, and open. Cinevva sits on the browser-first side of that convergence. Our bet is that game production collapses into a browser, ships instantly, and runs on hardware people already have. When neural simulation gets cheap enough to serve, it will slot into exactly that kind of distribution long before it slots into a 100 GB install. The scarce input doesn't change either way: something worth playing.

References

• MIRA live demo
• GitHub: mira-wm/mira
• Hugging Face: kyutai/rocket-science dataset
• arXiv: Multiplayer Interactive World Models with Representation Autoencoders
• Neurohive: MIRA, a world model fully simulates Rocket League

Tencent released Hunyuan Hy3 on July 6, and the headline isn't the benchmarks. It's the license. The full model, a 295-billion-parameter Mixture-of-Experts with 21 billion active per token, shipped under Apache 2.0 with no field-of-use or geographic restrictions. The April preview came with a restrictive license, and Tencent reversed that for the final release. You can now take a frontier-class Chinese reasoning model, put it in a commercial product, and owe nobody anything.

What Tencent actually shipped

Hy3 is a hybrid fast-and-slow-thinking model. It answers simple queries directly and switches into extended reasoning when the problem needs it, all in one system rather than two model variants. Context runs to 256K tokens. The weights are up on Hugging Face and ModelScope, with a rollout underway across OpenRouter and a stack of coding tools like Cline and Cherry Studio.

The pricing is the other half of the story. Through Tencent Cloud's TokenHub API, input tokens cost 1 yuan per million, about 15 US cents, with output at 4 yuan and cached input at a quarter of a yuan. Tencent says the model hits a 90% task-completion rate on agent workloads across its internal applications, and it has already been wired into Yuanbao, WeChat, ima, and the WorkBuddy coding assistant. Independent testing puts it ahead of other open models on agentic search and tool use, though it trails competitors like GLM-5.2 on some coding benchmarks, so it's a strong release rather than a clean sweep.

The Path of Exile deployment

Here's the part that matters for games. According to Tencent's announcement, the AI player assistant in Path of Exile: Advent on WeGame now runs on Hy3. The upgrade makes the assistant respond to player requests more accurately while cutting hallucinations, which is exactly the failure mode that has kept in-game AI helpers feeling like a gimmick. Path of Exile is a game whose build system is so deep that players keep wikis open on a second monitor. An assistant that actually understands the question is a real feature there, not a demo.

That's a concrete answer to a question the industry has been circling for two years: what does a language model actually do inside a shipped game? Not generate the game. Sit inside it, know its systems, and help the player. At 1 yuan per million input tokens, the economics of that feature just changed. An assistant that would have been a meaningful line item on a live-service budget a year ago is now cheap enough to be table stakes.

What it means for creators

We covered Tencent's Hunyuan 3D engine when it went global last November, and the pattern is the same one repeating a level up. First the asset pipeline commoditized. Now the intelligence layer is doing it too. Apache 2.0 weights at this quality mean any studio, any tool, any solo developer can run a frontier reasoning model on their own terms, fine-tune it on their game's systems, and ship an NPC or an assistant without a per-seat license or an API dependency they don't control.

For Cinevva this is the tailwind we build against. Every drop in the cost of intelligence makes AI features inside games, smart NPCs, in-game helpers, dynamic dialogue, less of an AAA luxury and more of a default. The gap between what a big studio can afford to put in a game and what you can is closing from both ends. Hy3 just closed it a bit more.

References

• Tencent: Hunyuan officially releases Hy3, advancing agent capabilities and deeper product integration
• TechNode: Tencent launches Hunyuan Hy3, integrates model across multiple products
• Caixin Global: Tencent launches final Hunyuan 3 model with free AI-agent feature
• Digital Applied: Tencent Hunyuan Hy3, an open-weight reasoning model
• Hugging Face: tencent/Hy3

Phantom Blade Zero has a date. S-GAME's dark wuxia action RPG, built on Unreal Engine 5, is now set for October 29, 2026, on PC and PlayStation 5, pushed back from a September target to buy finishing time. It crossed a million wishlists within 15 days of going up on Steam in December 2025, and Feng Ji, the producer of Black Myth: Wukong, shared its gameplay to his own following. When the person who made the biggest Chinese game in history vouches for the next one, people notice.

Phantom Blade Zero release-date trailer, The Game Awards 2025

One hit is a fluke, a wave is a shift

Black Myth: Wukong could have been a fluke. A singular team, a singular mythology, a singular moment. The test of whether China's single-player AAA moment was real was always going to be the second act: does anyone follow, and does the audience show up for a game without the Journey to the West brand behind it? A million wishlists in two weeks for a new IP from a smaller studio is a clear yes. The Chinese action-game push has stopped being a viral phenomenon and started being an export category with global attention, and Phantom Blade Zero is the flag-bearer of the follow-through.

The shape of it is worth naming. Wuxia and kungfu combat, steampunk edges, Unreal 5 fidelity, a hard single-player structure. This isn't China copying the Western AAA template. It's China shipping games that could only come from China, at a budget and polish that used to be a Western monopoly, and finding a worldwide audience that specifically wants that.

What it means for builders

The uncomfortable part, if you're a legacy studio, is that the moat drained. Unreal 5, frontier AI tooling, a generative asset pipeline like Tencent's Hunyuan 3D, and a talent pool sharpened on a decade of the world's most demanding live-service games, all of that is now available to studios that a few years ago wouldn't have been in the conversation. The production capability that separated the haves from the have-nots is commoditizing, fast. Phantom Blade Zero is what a mid-sized studio can do when the tools stop being the bottleneck and the only remaining variable is whether you have something worth saying.

That's the thread running through this whole China arc, and it's the one Cinevva is built on. When the cost and difficulty of production collapse, the winners are decided by point of view, not by budget. Wukong and Phantom Blade Zero prove it at the top of the market: a specific cultural vision, executed with tools anyone can now reach, beats a bigger studio with nothing particular to say. We're pushing the same collapse down to the individual creator, where the AAA pipeline shrinks to a browser and a prompt. The China wave is the loud, high-budget version of a shift happening at every scale at once. The tools are equalizing. What you make with them is the whole game.

References

• TechNode: Phantom Blade Zero to launch on Sept. 9, 2026, S-GAME says at TGA 2025
• Global Times: No longer a sideshow, China's AAA action games command global attention
• GameSpot: Chinese martial-arts RPG Phantom Blade Zero gets September 2026 release date
• Wikipedia: Phantom Blade Zero

Microsoft's fiscal year ended yesterday, June 30, and that date was the one the whole industry was watching. New Xbox CEO Asha Sharma spent June telling staff a reset was coming, and the cuts were timed to land once the books closed. As of today, July 1, the layoffs are reportedly imminent or already underway. Microsoft hasn't confirmed a headcount. A figure of around 1,000 roles has circulated in coverage but stays unverified.

Coverage of the Xbox reset and the layoffs expected after the fiscal-year close

The memo

On June 10, Xbox Wire published what was effectively a copy of an email sent to Xbox employees worldwide, co-signed by Sharma and chief content officer Matt Booty. It framed the next 100 days as a reset and was unusually blunt about the numbers. Revenue is down by nearly half a billion dollars over five years. The division's profit margin sits around 3%. Sharma called the trajectory unsustainable and said the business had expanded too far and invested too broadly.

The reset isn't only headcount. Reporting points to marketing budgets being cut hard, partner contracts being terminated ahead of the layoffs, and the cuts reaching companies Microsoft doesn't even own. Sharma is also said to be weighing a cheaper console tier. The memo itself didn't name the layoffs directly. It described the weaknesses and the need to course-correct, and everyone read the rest.

This is the Bungie story, one rung up

A month ago we wrote that Bungie ended active development on Destiny 2 and lined up its third round of layoffs since Sony bought it. The shape there was a studio that bet everything on a small number of huge live-service products, then had no smaller bet to fall back on when one wobbled. Before that, the March layoffs ran the same pattern across the industry.

Xbox is that story one level up. A platform that spent $69B acquiring Activision Blizzard and built a strategy around a handful of enormous studios and tentpole exclusives is now discovering that the fixed cost of all of it doesn't move when the audience does. The 3% margin is the whole thesis in one number. When your cost base is that large and that rigid, a soft year doesn't trim the edges, it forces a reset.

The industry tally backs this up. More than 3,700 verified game-industry job losses have been recorded in 2026 so far, with Bungie and Epic among the hardest hit. Xbox's cuts will push that number meaningfully higher.

What it means for builders

Three things worth holding onto, the same three we keep coming back to.

The size of the bet is the risk. The problem at Xbox isn't that any one game failed. It's that the strategy had no shock absorbers. When you concentrate spending into a few nine- and ten-figure products, a single miss isn't a setback, it's a crisis, because there's nothing smaller to fall back on.

Fixed cost is structural, not a mistake someone made. Any model where you need thousands of people and years of runway to ship a single title is fragile by design. A 3% margin means there's no slack in the system. The downturn didn't create that fragility, it exposed it.

The audience didn't leave, it spread out. Players aren't gone. They're across dozens of games, more of them made by small teams shipping fast. The studios that capture them next will be the ones cheap enough to make that a single miss doesn't end the company.

For Cinevva, this is the argument the platform is built on. The future isn't fewer, bigger games carried by thousand-person org charts and billion-dollar acquisitions. It's many more games made by small teams and individuals, shipped instantly in the browser, where the cost of trying something is low enough that one miss doesn't trigger a reset memo. The economics forcing Xbox to cut are the economics we built the platform to route around.

References

• Bloomberg: Xbox plans significant layoffs as it transforms under new CEO Asha Sharma
• Variety: Layoffs coming at Xbox? New CEO outlines upcoming 'reset' in memo to staff
• Game Informer: Report, Xbox preparing for significant layoffs as new CEO Asha Sharma details 'Xbox reset'
• Level Up: Xbox could face major layoffs in July as Asha Sharma reveals revenue fell nearly $500 million in 5 years
• Storyboard18: Gaming industry logs over 3,700 layoffs in 2026; Bungie, Epic Games among hardest hit

ByteDance unveiled Seedance 2.5 on June 23 at its Volcano Engine FORCE conference in Beijing. On the spec sheet it's a big jump: a single continuous 30-second clip in native 4K with 10-bit color and synchronized audio, generated in one pass. But the spec that matters isn't the resolution or the length. It's that you can now direct the thing. Seedance 2.5 takes up to 50 multimodal references per generation, lets you edit a finished clip in place instead of regenerating it, and blocks out a scene in 3D before it commits to a render.

A hands-on look at what Seedance 2.5 changed

The numbers, briefly

The jump from Seedance 2.0 is real. Single clips go from a few seconds to 30 seconds continuous. Output goes from a 1080p ceiling to native 4K with 10-bit color, not an upscale. Audio is co-generated in the same latent pass as the video instead of bolted on afterward, so it actually syncs. And the reference budget goes from roughly a dozen inputs to 50 multimodal references, mixing images, clips, audio, and style. ByteDance reports about 20% better prompt adherence on top of all that.

Those are the capability gains. They're impressive, and a year ago each one would have been the whole story. They aren't anymore.

Control is the new frontier

For two years the AI video race was about capability. Whose model does 4K, whose syncs audio, whose holds a character across shots. That race mostly resolved. Kling 3.0 shipped 4K with audio in March, LTX 2.3 did it open-source, and the frontier models are close enough that the differences are arguments, not gaps. Then the conversation moved to cost, with distilled models cutting the price of a second of video by 10x.

Seedance 2.5 points at the next axis: control. Three features make the case. First, targeted editing. If a character's hair color is wrong in a 30-second take, you fix that region instead of regenerating the whole clip and losing the performance, the expression, and the lighting you liked. Second, 50 references. That's enough to lock a character's face, a set's look, a prop, and a motion style all at once, which is how you get consistency across shots instead of a slot machine. Third, a 3D white-model previz step that lets you block out scene layout, camera framing, and motion before you spend a single full render.

Put together, that's a model you can direct, not just prompt. The shift from "can it generate something beautiful" to "can you get the exact thing you meant on the second try" is the one that turns these tools from demo machines into production tools.

What it means for builders

If you generate video inside a product, three things follow from this release.

Re-roll cost was the silent tax on every generative video feature. When the only way to fix one wrong detail was to regenerate the entire clip and gamble on getting everything else right again, users burned credits and patience in equal measure. In-place editing attacks that directly. The unit you pay to change shrinks from a whole clip to a region.

References are how you ship consistency. A character that looks the same across ten shots, a set that holds its look, a prop that stays the same prop. Fifty reference slots is enough to actually pin those down, and consistency is what separates a usable asset pipeline from a novelty.

Previz before render is the same lesson the cost story taught, from a different direction. Blocking a scene in cheap 3D before committing to an expensive 4K pass means you spend the costly compute on takes you've already framed, not on lottery tickets.

For Cinevva, this is the pattern we keep betting on. The generation tools in the platform only work as a business if creators can get what they meant quickly and cheaply, without re-rolling a dozen times to fix one thing. Capability got commoditized fast. Cost is dropping. Control, getting the exact result on purpose, is where the next round of useful products will be won.

Seedance 2.5 is in global enterprise beta now, with a wider public launch targeted for early July. No US availability timeline yet.

References

• The Next Web: ByteDance unveils Seedance 2.5, a 30-second native 4K AI video model that accepts 50 reference inputs
• Digital Applied: Seedance 2.5, ByteDance's 30-second AI video model
• MindStudio: Seedance 2.5, 30-second video, 4K, and 50 multimodal references explained
• Kie.ai: Seedance 2.5 release, what ByteDance just shipped

Godot 4.7 is out. The stable build landed on June 19, a week after RC2 cleaned up the last regressions, and right on the schedule the quiet release candidates pointed to. The team is calling it the "Director's Cut," and the headline is the same one we flagged at beta: more than 1,265 fixes from 309 contributors merged since 4.6 stable, with two rendering features and a rebuilt store doing most of the heavy lifting.

A rundown of what's new in the Godot 4.7 stable release

HDR output, finally on the panel

Godot has rendered internally in HDR for years, then tonemapped everything down to 8-bit SDR at the final output stage. 4.7 closes that gap. The engine now drives HDR displays directly on Windows, macOS, Linux via Wayland, iOS, and visionOS, for both 2D and 3D. Bloom, emissive surfaces, and wide-luminance scenes land on the monitor the way the renderer always intended, instead of getting crushed back to standard range at the last step.

The other rendering addition is the new AreaLight3D node, a rectangular light source with correct soft shadowing and distance falloff. Before this, you faked a ceiling panel or a window by scattering point lights and tuning attenuation until it looked passable. Now it's one node with a width and a height. Interior lighting setups that used to eat an afternoon take a minute.

A rebuilt Asset Store

The old Asset Library got replaced with a new Asset Store. It's not just a rename. Browsing is faster, previews are zoomable, and the whole thing runs on a background thread so loading assets doesn't freeze the editor while you work. For an engine whose ecosystem leans heavily on community add-ons, the store being pleasant to use is more than cosmetic.

The release rounds out with inline shader previews for editing materials in real time, transformable Control nodes for UI animation, a DrawableTexture2D for drawing straight into a texture, a MeshLibrary editor, and a stable Godot Android Build Environment with full Gradle support for building and publishing Android and XR projects. On the input side, 4.7 adds a built-in virtual joystick, gyro aiming, and accelerometer support, which matters for anyone shipping to phones.

The line item we watch: wasm64 web exports

C# hot-reload by default is a real quality-of-life win, but the feature we track most closely is the web export, and here the honest story has two halves. Web export in 4.7 still runs on WebGL 2.0 through the Compatibility renderer. The docs are blunt about it: Godot doesn't support WebGPU yet, and that's the prerequisite for running Forward+ or Mobile in a browser. What 4.7 does ship for the web is wasm64 exports, which lift the old 4 GB memory ceiling on browser builds.

That ceiling was a real wall. Big scenes, heavy audio banks, and large streamed worlds just didn't fit in a 32-bit address space, so teams cut content or skipped the web target entirely. With wasm64, a browser build can finally scale with the project. The rendering side is still the gap to close, something we wrote about back in January. With WebGPU now baseline in every major browser, the platform is ready and waiting. Godot's web renderer hasn't caught up to it yet.

For Cinevva, where games run instantly in a browser tab with no install, every step that makes engine web exports more capable is direct upside. It means more games that look good in a tab without a download, which is the bet the whole platform is built on.

Correction (July 14, 2026): an earlier version of this article said Godot 4.7 shipped WebGPU web export behind a flag. It didn't. Web export in 4.7 is still WebGL 2.0 only, and 4.7's real web addition is wasm64 exports. We've corrected the details above.

Should you upgrade

If you're on 4.6 and in active development, the migration is the usual one: open in 4.7, clear the deprecated API warnings, run your test scenes. The API stability promise tightened around 4.5 and has held, so compatibility breaks are rare. The RC cycle being as quiet as it was is the signal that this is a safe upgrade, not a risky one.

If you ship on Android or XR, the stable GABE release alone is worth the move. If you build for the browser, try a wasm64 export in a test branch and see how much headroom your project gains. Rendering on the web stays WebGL 2.0 for now, so keep your expectations there until Godot lands actual WebGPU support.

References

• Godot blog: Godot 4.7 is here
• 80.lv: Godot 4.7 has been released
• Phoronix: Godot 4.7 released with HDR output support
• GameFromScratch: Godot 4.7 released
• Coding Quests: Godot 4.7 everything new

Avataar.ai unveiled Varya at an event in New Delhi today, with the Secretary of India's Ministry of Electronics and IT in the room. It's a 14-billion-parameter AI video model that the company puts in the same quality bracket as Alibaba's Wan 2.2 and Google's Veo 3, but the number that matters isn't a quality benchmark. It's the price. Varya generates video at roughly ₹0.48 per second, which the company says is up to ten times cheaper than the leading global models. It does that by running a distilled architecture that cuts generation from the usual 50 diffusion steps down to four, while keeping output quality close to the full-step models.

Varya, India's distilled video model, generates clips in four steps instead of fifty

The model got there with help from India's IndiaAI Mission, which provided subsidized access to national AI compute. It's trained for Indian contexts on purpose, so it represents regional diversity, festivals, food, clothing, and everyday environments better than models trained mostly on Western data. There's a live demo at varya.avataar.ai, and Avataar says it will release Varya as an open-weight model, with training data, on India's AIKosh portal, so developers can self-host or fine-tune it. A technical report on the architecture and distillation method is coming too.

Four steps instead of fifty

The technical move here is the whole point. Diffusion models generate by starting from noise and refining over many steps, and each step is a full pass through the network. Fifty steps means fifty passes. Distillation trains a smaller, faster student model to reproduce what the slow many-step teacher produces, but in a handful of steps. Four steps instead of fifty is more than ten times less compute per frame, and compute per frame is where almost all the cost of video generation lives.

The wall-clock numbers make it concrete. On an NVIDIA H200, Avataar says Varya turns out a 5-second 720p clip in about 45 seconds. It clocks Wan 2.2 at around 1,230 seconds for the same job. That's not a small win at the margin, it's a different order of magnitude, and it's the difference between a model you can serve to a crowd and one you can only afford to run for a demo.

This is the same trick that made fast image generation practical over the last two years. Watching it land on video at production scale, with a real per-second price attached, is the signal worth paying attention to.

Why cost is the 2026 video story

For two years the AI video conversation was about capability. Whose model does 4K. Whose model syncs audio. Whose model holds a character across shots. That race mostly resolved. Kling 3.0 shipped 4K with audio in March, LTX 2.3 did it open-source, and the frontier models are all close enough that the differences are arguments, not gaps.

What didn't resolve was the economics. We wrote about how Sora burned roughly a million dollars a day against two million in lifetime revenue before OpenAI shut it down. The capability was never the problem. The cost of serving it was. A model that's beautiful and unaffordable is a demo, not a product.

So the interesting frontier in 2026 isn't a model that does something new. It's a model that does the same thing for a tenth of the cost. Varya is one data point. The broader move across the whole field, distillation, fewer steps, smaller models tuned for specific contexts, is the thing that turns AI video from a money-losing showcase into something a small team can actually build on.

What it means for builders

If you're building anything that generates video as part of the product, not as a one-off, the per-second cost is your unit economics. It's the difference between a feature you can offer everyone and a feature you ration or charge a premium for. A 10x cost drop doesn't make your product 10x better. It makes a whole category of products that were underwater suddenly viable.

This is why we keep watching the cheap end of the curve more than the flashy end. At Cinevva, the generation tools in the platform only work as a business if the cost of a generation is low enough that creators can experiment freely without us bleeding money on every click. The lesson from Sora was that great output at the wrong cost ends the company. The lesson from Varya is that the cost is finally starting to move, fast, and the teams that build on the cheap-and-good models will outlast the ones chasing the expensive-and-spectacular ones.

The regional angle matters too. A model trained for Indian contexts and priced for Indian budgets is a reminder that the next wave of these tools won't all come from the same handful of US labs, and won't all be priced for US wallets. Cheaper, more local, more specific is a different competitive axis than bigger and more general, and it's one the incumbents are slow to defend.

References

• TechCrunch: Cheaper, faster, and culturally aware, Avataar's video AI is built for India's scale
• Outlook Business: Avataar launches Varya, India's first distilled video AI model under IndiaAI Mission
• Republic World: This new India-built AI model wants to make video generation far more affordable
• WaveSpeed: AI video generation news 2026, latest models and updates
• MagicHour: AI video model release tracker 2026

Godot 4.7 reached its second release candidate today. The release has been baking since April 24, when beta 1 went out with 1,265 fixes from 309 contributors against 4.6-stable. RC1 landed on June 8 with just 16 fixes from 16 contributors, the kind of quiet snapshot that means the hard work is done. RC2 today mops up a few late regressions, and that's usually the last stop before a final build.

A hands-on overview of what's new in Godot 4.7

What RC2 actually fixed

RC2 is regression cleanup, not new features. The team caught a deadlock in WorkerThreadPool that could freeze multithreaded projects, and a related bug where the message queue stopped flushing before a MainLoop iteration finished. Both are the kind of intermittent hang that's miserable to reproduce and worse to ship. The rest of the list is editor polish: SceneTreeEditor drag operations that failed silently, AnimationTree crashes from a missing null check on root animation nodes, and a tilemap randomization bug that skipped the first tile.

None of these are glamorous. All of them are the reason you wait for a release candidate to settle before you move a production project over.

The features that ship with 4.7

We covered the headline features when beta 1 dropped, and nothing has changed since the feature freeze. The two worth repeating for anyone deciding whether to upgrade.

HDR output now works on every desktop platform. Godot has rendered internally in HDR for years, but the output stage tonemapped everything down to 8-bit SDR at the last step. 4.7 drives HDR monitors directly on Windows, macOS, Linux, and BSD, so bloom, emissive surfaces, and wide-luminance scenes finally land on the panel the way the renderer intended.

The new AreaLight3D node renders soft light from a rectangular surface. Before, you faked a ceiling panel or a window by scattering point lights and tuning attenuation until it looked passable. Now it's one node with a width and a height, with correct soft shadowing and distance falloff. Interior lighting setups that used to take an afternoon take a minute.

The release also brings C# hot-reload by default, an animation tree editor with proper undo/redo, and wasm64 web exports, which is the part we watch most closely.

What it means for web games

The wasm64 export is the line item that matters for browser games. Godot's web export still renders on WebGL 2.0 through the Compatibility renderer, and WebGPU support isn't there yet, so Forward+ and Mobile still can't run in a browser. What 4.7 changes is the memory side: wasm64 exports lift the old 4 GB ceiling, so big projects no longer have to cut content to fit a web build. The rendering gap is still the one to close, and we wrote about why in January.

For a platform like ours, where games run instantly in the browser with no install, an engine making its web target more capable is direct upside. More headroom for browser builds means more games that run well in a tab without a download, which is the bet Cinevva is built on.

Should you upgrade

If you're on 4.6 and your project is in active development, install RC2 in a branch and test now. The team's confidence in 4.7's compatibility is high, partly because the API stability promise tightened around 4.5 and has held. The migration path is the usual one: open in 4.7, clear the deprecated API warnings, run your test scenes.

If you ship on a fixed schedule, wait the few extra days for stable. RC2 being this quiet is the signal that the wait won't be long. The last beta-to-stable cycle for 4.6 went smoothly, and 4.7 looks set to repeat it.

References

• Godot blog: Release candidate Godot 4.7 RC1
• LinuxCompatible: Godot 4.7 RC2 fixes critical crashes and editor bugs before stable release
• Ziva: Godot 4.7 Beta features, HDR, release date and changes
• 80.lv: Godot 4.7's final development snapshot arrived

Summer Game Fest 2026 ran June 5 from the Dolby Theatre in Los Angeles, with Geoff Keighley and Lucy James hosting two hours of trailers. The headline slots went where they always go now. Square Enix announced Final Fantasy VII Revelation, the final chapter of the remake trilogy. Capcom revealed a Resident Evil Code Veronica remake for 2027. ArenaNet announced Guild Wars 3, beta in 2027. Alien: Isolation 2 showed up, PlatinumGames is making a Teenage Mutant Ninja Turtles game, and The Wolf Among Us 2 finally has a 2027 date.

It's a strong lineup if you like sequels and remakes, which is most of what the big show was. The more interesting signal came from the smaller stage running alongside it.

The full Summer Game Fest 2026 showcase

Day of the Devs is where indie lives now

The Day of the Devs Summer Game Fest Edition is the indie counterpart to the main show, and this year it had its usual run of weird, specific, personal games that the big broadcast doesn't have room for. New entries in the Trine and Yooka-Laylee series sat next to a long list of first-time studios. Over 1,700 games were submitted to be considered for the showcase, which is the real headline: the funnel into indie is enormous, and the screen time is tiny.

That ratio is the whole problem indie developers face, and it's the same one we keep circling back to. We wrote about Steam Next Fest February data showing prior momentum was the strongest predictor of success, and about the Switch 2 Indie World lineup earlier this year. A curated showcase like Day of the Devs is great if you're one of the twenty games picked. For the other 1,680, discovery is still a wall.

The AI line in the sand

The detail that stuck with me from the indie side: Saibot Studios went out of its way to say its horror game Tenebris Somnia used no generative AI, leaning on practical effects for its live-action scares. That's not a neutral production note anymore. It's a marketing position.

It lines up with the GDC 2026 State of the Game Industry survey, where 52% of developers said generative AI is having a negative impact on the industry, up from 30% the year before and 18% the year before that. The opposition is strongest exactly where you'd expect, among visual artists, narrative designers, and programmers, the people whose work the tools target most directly. The concerns that come up are consistent: training data and copyright, job displacement, and energy use.

So you've got a real and growing split. A chunk of developers and players now treat "no AI" as a quality and ethics signal, the same way "handmade" works for physical goods. Another chunk is shipping faster than ever by building with these tools. Both groups are at the same festival.

What it means for builders

We sit on the AI side of that line, and the survey numbers are worth taking seriously rather than waving off. The objections aren't going away, and "no generative AI" as a selling point is going to get more common, not less. That's fine. A studio choosing practical effects over generated ones is making a craft decision, and craft decisions are good for games.

The argument we'd actually make is narrower than "AI is good." It's that the tools change the economics of trying. The thing Day of the Devs proves every year is that there are far more people with a game idea than there are people who can afford the months it takes to find out if the idea works. Lowering the cost of getting from idea to playable prototype means more of those ideas get tested, and more weird, specific, personal games get made. That's the same outcome the no-AI indie crowd wants, reached by a different road.

The 1,700 submissions and twenty slots is the real constraint, and it's a discovery problem, not a tools problem. Whether a game was made with AI or with practical effects, it still has to be found. That's the part Cinevva is built to fix, and it's the part that no showcase, however good, solves at the scale the submission numbers demand.

References

• Summer Game Fest 2026 official site
• VGC: Summer Game Fest 2026, every announcement from the big live show
• GameSpot: Day of the Devs Summer Game Fest Edition 2026, everything revealed
• Game Developer: One third of game workers use generative AI, but half think it's bad for the industry
• GDC: 2026 State of the Game Industry report

Bungie is winding down the game that defined it. Jason Schreier reported for Bloomberg on May 21 that the studio is ending active development on Destiny 2, with a "significant" round of layoffs to follow. The June 9 update, Monument of Triumph, will be the final content drop. The servers stay up, but the development team behind nine years of Destiny does not have a greenlit follow-up, and there's no Destiny 3 in active production.

A breakdown of the Bloomberg report: development ends, layoffs are coming, and Destiny 3 isn't being built

What's actually happening

The decision is a hard stop on new development, not a shutdown. Destiny 2 stays online and playable. What's ending is the team that ships seasons, expansions, and the live-service treadmill that kept the game running since 2017.

These would be Bungie's third round of layoffs since Sony acquired the studio for $3.6B in 2022. The earlier rounds came after revenue targets slipped and the original sale terms tied a chunk of payout to performance Bungie never hit. This round is different in kind. The first two were cuts to keep Destiny 2 going. This one is cutting the people who made it because the game itself is being put into maintenance.

The resources are moving to Marathon, Bungie's extraction shooter and its bet on the next decade. That's the part that should worry anyone watching, because Marathon is not doing well.

The Marathon problem

Marathon launched and then fell off a cliff. By mid-April, a month after release, its 24-hour concurrent peak was hovering around 31,000 players, down from an all-time peak of 88,337 on launch day. That's a 59% drop from the high in about four weeks. For a live-service game built to run for years, losing more than half your concurrent players inside the first month is the kind of curve that ends projects.

So the math Bungie is running is brutal. The proven game with a loyal audience is being sunset to feed a new game that is already shedding players faster than it can be patched. Leadership is betting the studio's future on the title that's struggling, and paying for it by ending the one that worked.

Monument of Triumph: the send-off

The final update, Monument of Triumph, lands June 9 as a free download for every player. Bungie is treating it as a victory lap rather than a quiet patch. It brings back Sparrow Racing League as a permanent mode, reworks raid and dungeon loot, adds a world event, and ships a rewards pass with a new exotic hand cannon. For a game getting its last update, it's a generous one.

What's in the final Destiny 2 update, Monument of Triumph, arriving June 9

Why this keeps happening to live service

This is the same shape we wrote about with the March layoffs and, in a different industry, the Sora shutdown. A small number of huge, expensive products carry enormous fixed costs, and when the audience moves, the cost structure can't move with it.

A live-service game like Destiny 2 needs a large team shipping content forever just to stand still. The moment growth stalls, every season costs more than it returns, and the only lever left is to cut the team. The same dynamic that killed Sora (great launch, costs that scale with usage, an audience that didn't stick) is killing live-service game development from the other direction. The product worked, the costs didn't.

What this means for builders

Three things are worth holding onto here.

First, the size of the bet is the risk. Bungie's problem isn't that Destiny 2 failed. It ran for nine years. The problem is that the only follow-up was another nine-figure live-service game, so when that one wobbled, there was no smaller bet to fall back on. Studios that put everything into one giant title have no shock absorbers.

Second, the fixed-cost trap is structural, not a Bungie mistake. Any model where you need hundreds of people shipping content indefinitely to keep a game alive is fragile by design. The games that survive downturns are the ones where the cost of making and maintaining them is low enough that a soft launch isn't fatal.

Third, the audience didn't disappear, it fragmented. Destiny players aren't gone. They're spread across dozens of games, and the ones that capture them next will be cheaper to make, faster to ship, and easier to iterate on than a decade-long live-service monolith.

For Cinevva, this is the same argument we keep making. The future isn't fewer, bigger games carried by thousand-person teams. It's many more games made by small teams and individuals, shipped instantly in the browser, where the cost of trying something is low enough that one miss doesn't end a studio. The economics that just ended Destiny 2's development are the economics we built the platform to avoid.

References

• Bloomberg: Bungie Plans Layoffs After Ending Destiny 2 Development
• TechTimes: Bungie Confirms Third Layoff Wave, Destiny 3 Ungreenlighted, Destiny 2 Ends June 9
• Shacknews: Destiny 2 developer Bungie facing more layoffs
• The Game Post: Bungie planning significant layoffs, no Destiny 3 in development
• GameRant: Every Video Game Layoff and Studio Closure in 2026 So Far

Godot 4.7 beta 3 dropped today, focused on squashing regressions before stable. The release has been baking since April 24, when beta 1 went out with 1,265 fixes from 309 contributors against 4.6-stable. The headline features are HDR output across every desktop platform, a new AreaLight3D node, and a long list of editor and 3D pipeline improvements.

An overview of what changed in Godot 4.7 beta

HDR output, finally

Godot has rendered scenes internally in HDR for years. The piece that was missing was the output. 4.7 ships HDR output on every desktop platform: Windows, macOS, Linux, and BSD all now drive HDR monitors directly. The internal HDR pipeline can finally land its full dynamic range on the panel instead of getting tonemapped down to 8-bit SDR at the last step.

This is a much bigger deal than it sounds. Bloom, emissive surfaces, sun-disc highlights, and any scene with a wide luminance range now look the way the renderer always wanted them to. Combined with the existing tonemapper choices, it puts Godot's output in the same conversation as Unreal and Unity on consumer HDR hardware.

AreaLight3D

The new AreaLight3D node renders real-time light from a rectangular surface. Before 4.7, you'd scatter several point lights and tune attenuation to fake a soft area source. That's gone. You drop one node, set width and height, and get correct soft shadowing and proper distance falloff.

This is the kind of feature that makes interior lighting setups dramatically less painful. Studio softboxes, light panels, garage doors with daylight coming in: all of these were either ugly approximations or expensive light probe bakes. Now they're a single node.

Real HDR output in Godot 4.7

Ray-tracing polish and editor improvements

The ray-tracing pipeline introduced in 4.6 picks up meaningful fixes. Denoising quality is improved for static geometry. The probe placement tool got faster and more predictable. SDFGI's flicker on moving lights is reduced.

On the editor side, the 3D viewport got the usual round of quality-of-life: better gizmo precision, snapping improvements, and a new VirtualJoystick node that's useful for mobile and accessibility testing without leaving the desktop editor.

What else is in there

The release notes run long, but the items worth flagging for game devs:

• C# scripts now get hot-reload by default in the editor, which closes a long-standing gap with GDScript.
• The animation tree editor got an undo/redo overhaul that finally handles complex graph edits without breaking state.
• Web export gains wasm64 builds, lifting the old 4 GB memory ceiling. Rendering in the browser stays WebGL 2.0 for now, since the engine doesn't support WebGPU yet.
• Several Vulkan validation issues that caused intermittent crashes on Linux + Mesa were resolved.

Timing and stable release

Stable is expected late June. Beta 3 today is about regression cleanup, not new features, so this is the version to install if you want to test against your project before the final release lands.

For projects on 4.6, the upgrade path is the usual story: open in 4.7, hit the migration prompt, fix any deprecated API warnings, test. The team is unusually confident about 4.7's compatibility, in part because Godot's API stability promise tightened around 4.5 and has held.

This release follows the controversy in March about AI-generated pull requests overwhelming the maintainers. The fact that 309 human contributors landed 1,265 fixes since 4.6 suggests the project is still able to ship at pace despite the noise.

What we'd test first

For studios on 4.6, the HDR output is the single most visible upgrade. Open your most lit-up scene, switch to an HDR display target, and the answer to "should we ship 4.7" will be obvious within ninety seconds.

For new projects, AreaLight3D is the one to learn first. Build your standard lighting kit (sun, ambient, key, fill, kick) around area lights instead of points, and your scenes will look better than they did on 4.6 with zero extra effort.

References

• Godot blog: 4.7 beta 1 announcement
• Phoronix: Godot 4.7 beta with HDR output
• GameFromScratch: Godot 4.7 Beta released
• Ziva: HDR, AreaLight3D, and 1,265 Fixes

Google I/O kicked off today with the answer to the question everyone was asking: yes, there's a new video model, no, it's not called Veo 4. The model is Gemini Omni, and the rebrand is the strategy. Google is collapsing video, audio, image, and text generation into a single unified model under the Gemini name. Veo and Imagen as separate product lines are effectively folded in.

Demis Hassabis introduces Gemini Omni at the I/O 2026 keynote

What Gemini Omni actually is

Omni takes any combination of text, images, audio, and video as input and produces high-quality video as output, grounded in Gemini's world knowledge. The model edits through conversation: upload a clip, ask to rotate the framing, add an element, change the lighting, swap the actor's wardrobe. The pipeline is unified, which means the same model handles generation, editing, and continuation in one pass.

The day-one distribution is aggressive. Gemini Omni Flash is live now in the Gemini app, Google Flow, and YouTube Shorts for AI Plus, Pro, and Ultra subscribers globally. That's a much larger surface than any prior Google AI video launch, and it's a direct counter to the Sora shutdown three weeks ago. Google is making sure the audience Sora abandoned has an immediate Google-owned destination.

How it compares

Compared to Kling 3.0, Gemini Omni's headline differentiator is the multimodal input. Kling's strength is multi-shot continuity and 4K output; Omni's is the ability to mix references freely across modalities (a voice clip + a reference image + a prompt + a guide video, blended into one output).

The unified-model approach also changes the editing workflow. Where Kling and Runway treat editing as a separate model pass (typically slower and more expensive), Omni does edits in the same diffusion pass as generation. In practice this means iterations feel closer to text chat than to traditional video tools.

Pricing was not announced for the Pro tier on the keynote stage, but the Flash variant rolling out via the Gemini app means most subscribers get free access through their existing tier. That alone will pull a meaningful chunk of the AI video market into Google's gravity well.

Independent recap of the most striking Gemini Omni demos

Gemini 3.5 Flash and Google Antigravity

Two other announcements matter for builders. Gemini 3.5 Flash is the new mid-tier model, and it reportedly outperforms Gemini 3.1 Pro on coding and agentic benchmarks while running at Flash-tier speed and price. For any app calling Gemini today, the migration is a strict improvement.

Google Antigravity is Google's agent-first development platform. The pitch is that instead of building tools that help developers write code, Antigravity provides agents that help developers act, including across Google's product surfaces (Search, Workspace, Cloud, and the new Gemini Spark assistant). It's pitched directly against Anthropic's Claude with computer use and against OpenAI's GPT-based agent stack.

There was also the Android XR glasses preview, which is genuinely interesting but a year early to matter for game and creator workflows.

What this means for the AI video market

Three things shift today.

First, the "many specialized models" pattern that defined 2024 and 2025 (one for text, one for image, one for video, one for audio) is being explicitly killed by one of the labs that helped define it. Whether it sticks depends on whether Omni's unified output quality holds up across modalities in real use, not just in demos.

Second, the distribution moat just got bigger. Sora died because it couldn't get to a billion users fast enough. Omni starts with the Gemini app, Google Flow, and YouTube Shorts on day one. That's already where billions of people make and watch video.

Third, the open-source and price-competitive models (Kling at $0.07-0.14/sec, LTX-2 open-weights, Wan2.2) get an interesting position. If Google's bundled offer is "free with your $20/mo Gemini subscription," the only places to compete are either above (pro tools with deeper control) or below (open-source self-hosted).

What we'd test first

For anyone building creator tools, the immediate question is whether the Omni API ships at I/O developer day, what the pricing looks like, and what the rate limits are. The marketing surface (Gemini app, Flow, Shorts) is locked to Google. The API is the only way to integrate Omni into your own product. We'll update this post once those details ship.

For game studios specifically, the multi-shot consistency story is the one to watch. If Omni can hold a character across cuts the way Kling 3.0 can but with editable mid-stream control, cinematic prototyping for vertical-slice trailers gets a lot cheaper.

References

• Google: Introducing Gemini Omni
• Google: I/O 2026 keynote
• Google Developers: All the news from the I/O 2026 keynote
• Cybernews: Google pushes agentic AI at I/O 2026
• Tom's Guide: Biggest I/O 2026 announcements

Unity AI hit open beta today for every Unity 6 developer. It's the GDC preview, made real and shipped to the install base. The assistant lives inside the editor, understands your project, writes C# scripts, generates scenes from images, and creates placeholder assets without you ever leaving the dock.

Unity AI Open Beta launch trailer

Three modes: Ask, Agent, Plan

The assistant operates in three modes. Ask is the chat interface; you ask a question about your project and it answers using the codebase as context. Agent is the autonomous mode; you describe a goal, it generates a plan and starts editing scripts, prefabs, and scenes. Plan sits in the middle; the assistant proposes a plan, you approve it, then it executes.

It is built on third-party frontier models. Gemini is the announced backbone, but Unity is positioning the assistant as model-agnostic. This is not Unity Muse, the deprecated first-party-model product that came before it. The lineage matters: Unity tried to compete on raw model quality, didn't get there, and has pivoted to being the workflow layer.

AI Gateway and MCP server

Two pieces underneath the assistant matter more than the chat itself.

AI Gateway lets you connect third-party models (OpenAI, Anthropic, your own fine-tuned ones) through a single broker. Studios with security or licensing constraints can pin the assistant to a specific provider. The Gateway also handles credit accounting across providers.

Unity MCP server exposes the editor as a set of tools to any MCP-compatible client. That means Claude, Cursor, or your own scripts can drive the Unity editor from outside. Create a scene, place a prefab, write a script, run play mode, capture a screenshot, all over MCP. For pipeline automation and CI, this is the most consequential piece of the release.

Unity's official walkthrough of the MCP server

Pricing and access

Unity AI is a package you install into a Unity 6.0+ project linked to a Unity Cloud project. Personal users get 14 days free with 1,000 AI Credits, then $10/month for another 1,000 credits. Pro, Enterprise, and Industry seats include credits and MCP Server access by default. Credits are spent on model calls, with rates varying by the underlying provider.

The tiering is friendlier than the deprecated Muse, which gated more behind paid seats. For solo and indie developers, this is the first time Unity has shipped a serious AI assistant at a price that makes sense for hobby projects.

What works, what doesn't

Early reviews after the first few days line up with what you'd expect. Scaffolding new scripts, generating placeholder art, wiring up boilerplate, and explaining unfamiliar parts of the codebase work well. Anything that requires understanding the full state of a complex project (large prefab hierarchies, cross-scene references, performance-sensitive code paths) still needs hand-holding.

The Agent mode is impressive on small tasks and erratic on large ones, which matches the pattern of every coding agent shipping in 2026. The Plan mode (where you approve the work before it runs) is the safer default for anything non-trivial.

Why the timing is interesting

This ships eight weeks after GDC 2026's data showed that 36% of game developers are using genAI and over half think it's hurting the industry. Unity is betting that pulling the tools into the editor (rather than asking developers to context-switch to a chat app) shifts that ratio. If the assistant lives in the same window as the work, the friction of using it drops to near zero, and the holdouts have fewer reasons to opt out.

It also ships into a world where Unity games are heading into Fortnite under the Epic partnership. The "make a game" surface and the "ship to 500M Fortnite accounts" surface, in the same year, is a real go-to-market story.

What we'd test first

For studios already on Unity 6, the MCP server is the highest-leverage piece to wire up. Pair it with Claude Code or Cursor and you can drive the editor from your existing dev workflow. Asset import pipelines, automated playtest harnesses, and CI checks that need to open a scene all become scriptable from outside Unity for the first time.

For solo developers, Plan mode with the $10/month tier is a reasonable place to start. Use it for the parts of game dev that are obvious but tedious (UI screens, input handlers, save systems, settings menus) and reserve human time for game feel.

References

• Unity Discussions: Unity AI Open Beta is here
• PocketGamer.biz: Unity launches open beta
• 80.lv: Unity AI now in open beta
• Build Fast With AI: Getting started guide
• GamesBeat: Unity AI into open beta

Pika Labs released Pika Agents today, two days after Sora went dark. The product is a sharp pivot from "generate a clip" to "generate a presence." Each agent has a face, a voice, a personality, and persistent memory that survives across Slack, Telegram, Discord, X, Notion, Figma, and Google Meet. Once you've shaped one, it sticks around and learns how you like ideas framed.

A walkthrough of what Pika Agents can do on day one

PikaStream 1.0: the engine under the hood

The real-time video chat is powered by PikaStream 1.0, a model Pika first previewed on April 2. It generates personalized 24 FPS video at 480p on a single H100, with end-to-end speech-to-video latency around 1.5 seconds. The agent reacts with synchronized lip-sync, facial expressions, and what Pika calls "emotionally appropriate" body language. Pricing is $0.20 per minute of streamed video.

The integration model is the interesting part. Instead of building a Pika app, Pika ships skills. The Google Meet skill lets you invite your "AI Self" into a meeting as a participant. The agent joins the call, listens, and responds in video. Zoom and FaceTime are listed as next.

Memory that persists across surfaces

Pika is selling agents as a single entity that follows you across tools. The agent you've trained in Discord remembers your context when you ask it something in Notion. It remembers the running joke from yesterday and the brief you gave it last week. For repeat creative work (a weekly newsletter cover, recurring brand video, agency client whose voice you keep getting wrong) the persistence is the actual product.

Early users have been training agents on specific recurring tasks: storyboard-to-animation, brief-to-ad-cut, ongoing character consistency across episodic content. The agent owns the rhythm. You direct.

Pika 2026 feature breakdown including Agents and PikaStream

The launch campaign is its own thing

Pika commissioned a launch film with production company Ceiling Train and director Josh Cohen. The film leans hard into the "Black Mirror" framing, with users "birthing" their AI Selves and letting them loose. It trended on X for two days straight. Whether that's a feature or a warning depends on which side of the AI agent debate you sit on.

Why this matters for game and creator workflows

For game studios, persistent video agents are an obvious fit for NPC prototyping, voice direction, and table-read sessions. Drop the agent into a Discord channel for the writing team and treat it as the in-canon voice of a character you're developing. For creator workflows, the agency-team-of-one play is real: one operator, three agents trained on three clients, all running in parallel across Slack and Meet.

The unit economics also look durable in a way Sora's didn't. PikaStream's H100-per-stream cost at $0.20 per minute is roughly margin-positive at current GPU prices. The agent personality data, not the model weights, is where the moat lives.

What we'd test first

If you're a small team, the highest-value experiment is probably the Google Meet skill. Train an agent on your product positioning, invite it to your next external pitch as a silent participant, then have it summarize the conversation back to you. It's the AI meeting note-taker pattern but with a face that the other side can see and react to.

For larger studios, the more interesting test is multi-agent. Two agents in the same channel, each trained on different parts of a brand, debating a creative choice while a human moderates. That's the workflow shape that wasn't possible before persistent memory across surfaces.

References

• Pika: Introducing real-time video chat for agents
• PikaStream 1.0 launch details
• LBBOnline: Pika's AI Selves launch film
• Eyerys: Pika Labs launches Agents

OpenAI pulled the plug on Sora today. The web app and the mobile app are dark as of April 26, 2026, exactly 33 days after the March 24 announcement. The Sora API will follow on September 24. Video generation from text prompts has also been removed from ChatGPT. Six months after the splashiest AI video launch of the cycle, the product is dead.

Why OpenAI shut Sora down so fast, and what comes next

The numbers behind the decision

External reporting puts Sora's inference burn at roughly $1M per day, with peak-cycle estimates as high as $15M per day during launch week. Lifetime revenue came in at around $2.1M. That is not a rounding error you can fix with a price tweak.

Downloads collapsed alongside the costs. After the launch surge that made Sora the fastest-rising AI app of late 2025, monthly downloads fell from 3.3M to 1.1M, a 67% drop. Global active users dropped under 500,000 by the time of the shutdown. The product never found a second wave of users who would pay for credits at the rate the GPUs needed.

Why video AI is structurally harder than text AI

A second of video output costs one to two orders of magnitude more to generate than a paragraph of text. The cost curve does not bend the way LLM inference did, because video doubles work on two axes (frames and resolution) and the diffusion process can't take the same shortcuts that token streaming allows.

Sora's pricing tried to hide that with a credit model, but consumers found credits both expensive and confusing. Meanwhile Google's Veo 3.1 and Kuaishou's Kling 3.0 closed the quality gap inside eighteen months while running on better unit economics, in part because their parent companies have ad and short-video businesses that subsidize the compute.

The Disney deal that didn't survive

The collateral damage is the December 2025 Disney partnership: a $1B investment and a three-year character licensing deal for over 200 Disney, Marvel, Pixar, and Star Wars characters. Disney learned about the shutdown less than an hour before the public announcement. No money changed hands. The character licensing rights are now back with Disney, and the AI video studios still standing are unlikely to land anything close to that scale in the next year.

Where Sora users went next: Kling, Veo, Runway, and Pika

Where the users went

The week after the shutdown announcement, Kling AI saw weekly active users jump 4% to 2.6M. Runway and Vidu both reported similar bumps. Pika followed up with Pika Agents on April 28. The market did not contract. It re-allocated to operators with sustainable cost structures and to teams shipping new product surfaces rather than scaling the same monolithic generator.

What this means for builders

Three takeaways for anyone shipping generative tools.

First, the days of pricing AI video on "credits per generation" while quietly losing money are over. Either you have a host business that subsidizes the compute (Kuaishou, Google) or you charge for outcomes and own the workflow (post-production, marketing, game cutscenes).

Second, the moat is in the surface, not the model. Pika's bet on agents and Kling's bet on the 7-in-1 editor are both about owning the creative workflow end to end. The model is a component.

Third, lock-in matters less than it used to. Most professional shops were already running two or three video models in parallel for any serious shot. Sora's exit changes the rotation but not the workflow.

For Cinevva, this is good news. Our position has always been that the engine and the runtime matter more than any single model, and that creators want to mix outputs from many providers. Sora's collapse confirms the underlying economics that drove that decision.

References

• TechCrunch: Why OpenAI really shut down Sora
• CNN: OpenAI is shutting down its Sora video app
• OpenAI: What to know about the Sora discontinuation
• TechCrunch: Sora's shutdown could be a reality-check moment
• Bloomberg: Kling, Runway, Vidu set to replace Sora

Three.js r184 shipped this week. The headline is a new HTMLTexture class that lets you point Three.js at any live HTML element and use it as a texture on a mesh. Buttons stay clickable, inputs stay focusable, CSS animations keep ticking, and the browser handles accessibility and IME the way it always does.

Chrome's official walkthrough of the HTML-in-Canvas API that backs Three.js HTMLTexture

HTMLTexture, in one paragraph

HTMLTexture wraps the Chrome 146 HTML-in-Canvas API. You pass it a DOM element. Three.js sets layoutsubtree on the canvas, parents the element into the renderer's DOM, and projects its rasterized output into a THREE.Texture you can sample from any material on WebGL 2 or WebGPU. The official examples (webgl_materials_texture_html and webgpu_materials_texture_html) include formatted text, images, SVG, a working input, and a button.

In practice this means React components, design system widgets, video players, charts, or full embedded pages can sit on a curved screen, a billboard, or a UI panel inside a Three.js scene without the usual "rasterize to a canvas, then forget about interactivity" tradeoff.

The catch

The feature only lights up where the underlying API exists. Right now that is Chromium 146 and above with the chrome://flags/#canvas-draw-element flag flipped on, which means production sites still need a fallback. The PR keeps the door open by making HTMLTexture fail gracefully when the API is missing, but you will want to feature-detect before promising clients a fully interactive Three.js dashboard.

Safari and Firefox have not signaled timelines. Given that WebGPU itself only just reached baseline across all three browsers in late 2025, HTML-in-Canvas will likely take a similar two-year arc.

A community demo of HTML-in-Canvas before the Three.js wrapper landed

The other shoe: zero per-frame allocations

The less flashy but arguably more important change is in the renderer's hot path. Before r184, rendering 1,000 meshes at 60 FPS could allocate 240,000 to 500,000 throwaway objects per second. The garbage collector then had to deal with that on every long-running scene. r184 eliminates those allocations on the core render loop. Long-running WebXR sessions, dense construction viewers, and any scene that lives past the first 30 seconds gets a real frame-time win for free.

Smaller things worth flagging

AnimationAction now preserves interpolant settings on creation, so a timeScale reversal no longer jumps. AudioLoader no longer races the loading manager. BatchedMesh and InstancedMesh stop throwing from getColorAt when colors haven't been set. The deprecated WebGL instancing render paths are gone for good. Examples gained an SSGI ball pool, volumetric clouds on the 3D Tiles loader, and banking on the roller-coaster camera.

What we'd do with it

For Cinevva, the most interesting use case is on-canvas UI for the reels player and creator tools. Right now we composite HTML over Three.js. With HTMLTexture, the chat overlay, the asset browser, and the prompt input can all live on geometry inside the scene without losing keyboard navigation. As soon as the API ships unflagged in Chrome stable, we'll wire up a feature-flagged variant.

References

• Three.js r184 release notes
• HTMLTexture pull request
• Chrome for Developers: HTML-in-Canvas origin trial
• HTML-in-Canvas community demos

Babylon.js 9.0 is out, and it's the biggest release the open-source WebGPU/WebGL engine has shipped to date. The headline feature is a clustered lighting system that lets scenes render hundreds or thousands of dynamic lights at smooth frame rates on both WebGPU and WebGL 2.

The new Node Particle Editor in Babylon.js 9.0

Clustered lighting changes everything

Traditional per-pixel lighting forces every pixel to compute contributions from every light in the scene, even lights nowhere near it. Clustered lighting fixes that by grouping lights into screen-space tiles and depth slices. At render time, each pixel only evaluates lights that actually affect it. The result is that a scene with 1,000 dynamic lights can now run at real-time frame rates instead of crawling.

This works on WebGPU and WebGL 2, which means the performance improvement reaches the widest possible audience.

Textured area lights

Area lights were introduced in Babylon.js 8.0. Version 9.0 adds emission textures, so you can project any image through a rectangular area light. Think stained glass projections, LED panel displays, or cinematic lighting setups with physically accurate light emission.

Volumetric lighting

The new volumetric lighting system renders light shafts through fog, dust, and haze. It includes configurable extinction and phase parameters for artistic control over how light scatters in the atmosphere. The system supports directional light sources and uses WebGPU compute shaders where available, with WebGL 2 fallbacks.

Node Particle Editor

The Node Particle Editor (NPE) is a visual drag-and-connect tool for building particle systems. If you've used the Node Material Editor, the workflow is identical. You get full control over emission shapes, sprite sheets, update behaviors, and sub-emitters through a non-destructive node graph. Flow maps and gravity attractors round out the particle upgrades.

Gaussian splatting overhaul

Babylon.js introduced gaussian splatting support in version 7.0. Version 9.0 rewrites the system with shadow casting, triangle splatting for opaque mesh-like rendering, and support for compressed SOG/SOGS formats alongside .splat, .ply, and .spz files. You can now combine multiple splat assets in a single scene with global sorting, and each part can be independently transformed and animated. Adobe contributed to this work.

Frame Graph system

The Frame Graph, introduced as an alpha feature in 8.0, hits v1 in this release. It's a directed acyclic graph where each node represents a rendering task. The system manages texture allocation, reuse, and optimization automatically. The team reports GPU memory savings of 40% or more in some cases. You can compose custom render pipelines visually through the Node Render Graph Editor or programmatically through the class framework.

Other additions

Animation retargeting lets you apply animations from one character skeleton to another, even with different bone structures and naming conventions. The release also includes a rebuilt inspector, large world rendering to prevent floating-point jitter in large coordinate systems, and geospatial rendering capabilities.

References

• Announcing Babylon.js 9.0
• Babylon.js Forum
• Babylon.js Documentation

Kuaishou has rolled out Kling 3.0 globally after first announcing it on February 4. The update brings native 4K/HDR output, synchronized audio with lip-sync, multi-shot narrative generation, and motion control that lets you extract movement from one video and apply it to entirely different characters.

Overview of Kling 3.0's new capabilities

4K with native audio

Kling 3.0 generates video at 4K resolution (3840x2160) and 30 FPS directly from the diffusion process. There's no post-generation upscaling. The audio generation is built into the same forward pass, including lip-synced dialogue in five languages: English, Chinese, Japanese, Korean, and Spanish. Previous versions required separate dubbing and audio sync steps. That's gone now.

Multi-shot narrative sequences

The standout feature is multi-shot generation. You can generate a series of up to 6 connected shots from a single session, with the model maintaining character consistency, lighting, and spatial continuity across cuts. Each shot can be up to 15 seconds.

This is aimed directly at short-form narrative content. Instead of generating individual clips and trying to stitch them together (and dealing with the inevitable character drift between generations), you describe a sequence and get back shots that look like they belong in the same scene.

Motion control and transfer

Kling 3.0 can extract motion from a reference video and apply it to new characters. You can take a dance sequence, a fight choreography, or a specific gesture from one video and map it onto a generated character. The model includes a motion brush for frame-level control and 6-axis camera controls for precise shot composition.

There's also an @-syntax for prompting that lets you reference specific characters or elements by name. If you've established a character in one generation, you can call them back in subsequent prompts.

Deep look at what Kling 3.0 changes for AI video

The 7-in-1 editor

Kling 3.0 ships with what Kuaishou calls a "7-in-1 multimodal editor." It consolidates text-to-video, image-to-video, video-to-video, motion transfer, lip-sync, extend, and retake into a single workspace. The goal is to eliminate the fragmented toolchain that AI video production typically requires, where you'd need separate tools for generation, dubbing, color matching, and shot assembly.

How it compares

Kling 3.0 competes directly with OpenAI's Sora 2 Pro and Google's Veo 3.1. On paper, Kling has the resolution advantage (native 4K vs. Sora's 1080p) and the price advantage ($0.07-0.14 per second vs. Sora's $0.10-0.50). It also offers a generous free tier of 66 credits per day, while Sora has none.

Sora 2 still leads on physics simulation, handling light refraction, fluid dynamics, and collision physics more convincingly. It also scores higher on baseline visual fidelity in side-by-side comparisons. The practical advice from most reviewers is to use both and pick the better output for each specific shot.

For game developers, Kling 3.0's multi-shot generation and motion transfer are the most interesting features. Being able to generate a consistent sequence of cinematic shots, with matching characters and lighting across cuts, is directly applicable to trailer production and cutscene prototyping.

References

• Financial Content: Kling 3.0 global release
• fal Academy: Kling 3.0 overview
• VO3 AI: Kling 3.0 vs Sora 2 Pro vs Veo 3.1

March 2026 might be the month that "vibe coding" crossed from novelty to production. Multiple solo developers and two-person teams shipped complete, playable games built primarily with AI tools. Not prototypes. Not demos. Actual finished products, some of them on Steam.

Void Balls: 29,000 lines in 10 days

BigDevSoon shipped "Void Balls," a bullet-hell platformer, in just 10 days. The stats are wild: 29,000 lines of C#, 173 scripts, 88 test files, 5 enemy types, and 15 power-ups. Their workflow used 8 parallel Claude Code agents, each handling a different domain of the game (physics, enemies, UI, audio, and so on).

Art came from Replicate models. They used Flux 2D Game Assets for sprites and Recraft V4 Pro for UI elements. Audio was generated through ElevenLabs. The whole thing was orchestrated by a single developer directing AI agents like a project manager running a team.

Building full games with Claude Code vs GPT-5 Codex

Grumbulus: two evenings, two developers

Two developers built "Grumbulus," a browser game where you play as a storm cloud destroying a city, in two evenings. The game runs on 15,000 lines of vanilla JavaScript with procedural audio and parallax rendering. All code was generated through Claude Code. No frameworks, no game engine. Just raw JS and a couple of people with a funny idea.

Weekend roguelike

Barret Blake built a web-based roguelike dungeon crawler over a single weekend using Claude Code. The game features procedural dungeon generation, turn-based combat, inventory management, and multiple enemy types. Blake documented the process on his blog, showing how he went from concept to playable game in roughly 48 hours.

Catvivors hits Steam

Przemek Chojecki, an AI researcher, released "Catvivors" on Steam Early Access on March 13. It's a roguelite auto-battler where you play as a cat fighting waves of enemies, built with Claude Code and Codex. The game has 6 playable characters, 15+ weapons with evolutions, and 24 hand-crafted levels. It's priced at $4.99 and runs on Windows, Mac, and Linux.

CODEX MORTIS: "100% AI-developed"

The boldest claim came from CODEX MORTIS, which launched on Steam Early Access on March 26 billing itself as the "world's first 100% AI-developed game." Developer GROLAF says every line of code, every sprite, and all music was created through AI tools.

The game itself is a necromantic survival bullet hell. A free demo in December 2025 attracted 10,500 players who averaged 71-minute sessions, over four times the industry average for indie demos. It landed on Steam's Popular Upcoming list before launch. Twitch streamer Asmongold played it and called it "pretty good."

Testing AI-powered game development with Cursor and Claude

What this means

The common thread across all these projects isn't that AI replaced the developer. It's that AI compressed the timeline. Tasks that used to take weeks now take days. A solo developer can now run multiple AI agents in parallel, effectively acting as a creative director rather than writing every line by hand.

The first 80% comes together fast. The last 20%, the polish and game feel, still needs human judgment. But the barrier to shipping a complete game has dropped dramatically.

References

• BigDevSoon: Building games with AI
• GonzoML: How we built Grumbulus
• Barret Blake: AI roguelike
• Przemek Chojecki: AI helped me solo dev a game on Steam
• GamesPress: CODEX MORTIS Early Access launch

Kickstarter had a remarkable March. A Cyberpunk-themed trading card game became the most-funded TCG in the platform's history. An MMO remake hit a percentage-funded record so absurd it looks like a typo. And a brand-new Metroidvania for the original NES funded in a single day. Here's what happened.

Cyberpunk TCG: $8.3M and the all-time TCG record

The Official Cyberpunk Trading Card Game, developed by WeirdCo in collaboration with CD Projekt Red, launched on Kickstarter on March 19 and hit its $100,000 goal in five minutes flat. Within 24 hours, it had raised over $8.3 million from more than 10,000 backers, making it the most-funded trading card game in Kickstarter history.

The game features unique mechanics including "gig dice" (ranging from D4 to D20) that determine street credit and outcomes, plus a "Legends" system with three face-down character cards. Backing tiers range from $49 for two starter decks up to nearly $8,000 for premium packages. WeirdCo says they already have over five years of new gameplay mechanics planned.

Fulfillment is scheduled in two waves: Q3 2026 for express shipping and Q4 2026 for standard.

Everything you need to know about the Cyberpunk TCG

AdventureQuest Worlds: Infinity: 100,000,000% funded

AdventureQuest Worlds: Infinity set a new Kickstarter percentage record by hitting 100,000,000% funded. That's one hundred million percent. The campaign reached that number in just 2.5 days after launching on March 6. All stretch goals were cleared in under an hour.

The project is a full remake of AdventureQuest Worlds, the browser-based MMORPG that's been running for 18 years. Developer Artix Entertainment is preserving all existing player accounts and character data in the transition. The new version runs at 60+ FPS with keyboard and gamepad support across Steam, phones, and tablets.

The community response reflects a deep attachment to the original game. Players who've invested years of progress aren't just funding a new game. They're funding the continuation of their characters.

The AQW Infinity Kickstarter that broke a world record

Annalog: a new NES Metroidvania

Annalog is a complete Metroidvania built for the original Nintendo Entertainment System. The Kickstarter launched on March 26 and hit its $7,000 goal within a day. The game is already 100% finished and tested on real NES hardware, with the source code freely available on GitHub.

The game features thirteen regions, eight bosses, dozens of characters, and three possible endings. Players reprogram abandoned machines to solve puzzles, and there are multiple paths through each area. The Kickstarter funds a physical cartridge release with a boxed copy and printed manual. Stretch goals include a Nintendo Switch port at $20,000 and a 1990s-style gaming magazine write-up at $15,000.

Foes of Middle-earth: $600K+ for a LOTR dungeon crawler

Foes of Middle-earth, a cooperative dungeon crawler board game set in Peter Jackson's Lord of the Rings universe, raised over $600,000. The game is officially licensed from New Line Cinema and features miniatures of Frodo, Aragorn, Legolas, and Gimli facing off against the Balrog, Nazgul, Orcs, and Wargs.

The campaign funded its $50,000 goal in 10 minutes. Up to four players explore randomized tile maps, level up heroes, and complete personal quests across 2-3 hour sessions. The base game costs $159.

References

• Polygon: Cyberpunk TCG raises $8M in 24 hours
• Player.one: Cyberpunk TCG most-funded TCG ever
• GamesPress: AQW Infinity 100,000,000% funded
• COGconnected: Annalog hits 100% in one day
• Polygon: Foes of Middle-earth raises $600K

Microsoft Research released TRELLIS.2, a 4-billion-parameter image-to-3D model that generates textured 3D assets with full PBR materials. It's MIT licensed, open source, and fast: 3 seconds at 512 cubed resolution, 17 seconds at 1024 cubed, and 60 seconds at 1536 cubed on an NVIDIA H100.

Microsoft's TRELLIS turning 2D images into fully textured 3D models

What it generates

Unlike earlier image-to-3D models that output basic meshes with baked vertex colors, TRELLIS.2 produces complete PBR material maps: albedo, metallic, roughness, normal, and opacity. These are the same maps that game engines and rendering pipelines expect, which means generated assets can go straight into Blender, Unreal Engine, or Unity without manual material setup.

The model uses flow-matching transformers and a novel O-Voxel representation. O-Voxel is a "field-free" sparse voxel structure with 16x spatial compression that handles arbitrary topology. That includes open surfaces, non-manifold geometry, and internal structures that previous iso-surface methods couldn't represent.

Speed and resolution

The generation speed scales with resolution:

Mesh-to-O-Voxel conversion takes under 10 seconds on CPU. O-Voxel back to mesh takes under 100 milliseconds with CUDA acceleration.

How it compares to TRELLIS 1

TRELLIS.2 doubles the parameter count from 2 billion to 4 billion. The original TRELLIS could generate decent geometry but lacked proper PBR materials. Version 2 adds the full material pipeline, handles more complex topology, and runs at higher resolutions.

Tutorial: creating 3D models from 2D images with TRELLIS

How to use it

There are several ways to run TRELLIS.2:

Hugging Face demo. A free web demo lets you upload an image and get a 3D model back without any local setup.

Local deployment. The GitHub repository includes Docker images for local installation. You'll need an NVIDIA GPU with at least 24GB of VRAM for comfortable operation.

ComfyUI integration. Community-built ComfyUI nodes let you incorporate TRELLIS.2 into existing generation workflows alongside other AI tools.

Why this matters

Open-source image-to-3D has been improving steadily, but most models produce either decent geometry with poor textures or good-looking results that fall apart when you inspect the topology. TRELLIS.2 addresses both by generating clean geometry with proper PBR materials at production-relevant resolutions. The MIT license means you can use it commercially without restrictions.

For game developers and 3D artists, this is a useful prototyping tool. Drop in a concept art image, get a textured 3D model in seconds, and refine from there.

References

• TRELLIS.2 project page
• TRELLIS.2 on GitHub
• 3dtrellis.com

March 2026 has brought another wave of job cuts across the game industry. Ubisoft gutted a studio co-founded by Tom Clancy, nDreams shuttered two VR studios, and EA trimmed the teams behind Battlefield 6 just months after a record-breaking launch. Meanwhile, a GDC survey paints a bleak picture of how widespread the damage has become.

Ubisoft Red Storm: 105 jobs gone

Ubisoft laid off 105 employees at Red Storm Entertainment, the North Carolina studio Tom Clancy co-founded 30 years ago. The studio has stopped all game development. About 70 staff remain to handle IT support and maintenance work on the Snowdrop engine.

The cuts are part of Ubisoft's broader restructuring into five "creative houses" organized by genre. The company's stock has lost over 80% of its value since 2021, and it announced plans to cut a total of 400 employees in 2026 across multiple rounds of layoffs. Studios in Leamington, Osaka, and Saint-Antoine were already closed in 2025.

Ubisoft's ongoing crisis explained

nDreams: 78 roles at risk

UK-based VR specialist nDreams put 78 employees at risk of redundancy on March 5. The company closed its Brighton-based Near Light and Compass studios entirely. The restructured company will focus on its Elevation studio with roughly 120 staff working on unannounced projects, plus a small XR R&D group.

CEO Tim Gillo pointed to "undeniable commercial challenges" in the VR games market. Compass had only been established in February 2025, making it barely a year old at the time of closure.

EA: Battlefield studios cut despite record launch

Electronic Arts cut staff across Criterion, DICE, Ripple Effect, and Motive, the four studios collectively working on Battlefield 6. EA called it a "realignment." The irony is hard to miss. Battlefield 6 was the best-selling game in the US in 2025 and moved roughly 7 million copies at launch in October.

But the game hasn't held its audience. Steam concurrent players dropped from a peak of 747,700 to around 67,000 by the time of the cuts. Player complaints about monetization, balance issues, and slow content updates appear to have caught up with the franchise.

IGN covers the Battlefield studio layoffs

The GDC survey numbers

The 2026 State of the Game Industry report surveyed over 2,300 professionals and the results are stark. 28% of respondents experienced a layoff in the past two years. That number rises to 33% for US-based developers. Game designers were hit hardest at a 20% layoff rate.

Of those laid off, 48% haven't found new work. AAA studio employees were disproportionately affected, with two-thirds reporting layoffs compared to one-third at indie studios. 74% of surveyed students expressed concern about their future job prospects.

On unionization, 82% of US respondents support it. Support ran highest among workers earning under $200,000 (87%), those who'd been laid off (88%), and those under 45 (86%). Not a single respondent aged 18-24 opposed unionization.

References

• PC Gamer: Ubisoft lays off 105 at Red Storm
• GamesIndustry.biz: nDreams restructuring
• Game Informer: Battlefield studios hit with layoffs
• GDC 2026 State of the Game Industry
• GamesIndustry.biz: GDC survey details

Three.js r183 shipped on February 18, and it's a bigger architectural shift than most point releases. The headline change: PostProcessing has been renamed to RenderPipeline. That's not just a rename. It reflects how the class evolved from simple post-processing management into a full pipeline system that handles multiple rendering passes and history buffers.

Introduction to Three.js WebGPU and the TSL shader node system

RenderPipeline replaces PostProcessing

The new RenderPipeline class is designed specifically for WebGPURenderer. It manages the full rendering pipeline setup, post-processing effect chains, and defines the final output through an outputNode. The outputColorTransform property controls default tone mapping and color space transformation.

This matters because the old PostProcessing name was misleading. With temporal anti-aliasing (TRAA) integration for volumetric lighting, screen space reflections with temporal stability, and HDR pipeline support using HalfFloatType output buffers, the class was already doing much more than post-processing. The new name matches reality.

Clock deprecated, Timer arrives

The Clock class is now deprecated in favor of a new Timer class. If you're using Clock in your animation loops, start migrating. The new Timer provides a cleaner API for the same time-tracking functionality.

BatchedMesh gets per-instance control

BatchedMesh gained per-instance opacity and wireframe material support. This is a significant addition for open world scenes and city builders where you need to fade individual instances in and out without creating separate materials. It also means you can selectively apply wireframe rendering to specific instances for debugging or stylistic effects.

WebGPU post-processing in Three.js with React Three Fiber

Material improvements

MeshLambertMaterial and MeshPhongMaterial now support scene.environment image-based lighting (IBL). Previously these cheaper materials couldn't use environment maps for ambient lighting, pushing developers toward the more expensive MeshStandardMaterial. Now you can use IBL with legacy materials where PBR accuracy isn't needed.

MeshPhysicalMaterial adds clearcoat support for rect area lights. If you're rendering car paint or polished surfaces lit by area lights, the clearcoat layer now responds correctly.

Performance optimizations

Matrix4.decompose() now caches the determinant calculation, and invert() has been optimized. These are called thousands of times per frame in scenes with many objects, so the accumulated savings add up.

TSL spec and clipSpace node

Three.js r183 introduces a formal TSL specification document (TSL.md) and a new clipSpace node. TSL (Three.js Shading Language) lets you write shader code in JavaScript that compiles to either WGSL for WebGPU or GLSL for WebGL. The new spec formalizes the node system's API, making it easier for third-party libraries to build on top of it.

The clipSpace node gives direct access to clip-space coordinates in TSL, useful for screen-space effects and custom projection techniques.

References

• Three.js r183 release notes
• RenderPipeline documentation
• TSL specification
• Three.js WebGPU migration guide

The itch.io desktop app shipped version 26.6.0, its first stable release in a while. The update modernizes the app's infrastructure with native Apple Silicon support, a new authentication flow, and the end of 32-bit builds.

Comparing Steam and itch.io as platforms for indie game distribution

Native Apple Silicon

macOS users on M-series chips will now get a native arm64 binary instead of running through Rosetta 2. The app auto-updates to the arm64 version, so existing users don't need to do anything manually. This should noticeably improve performance and battery life on Apple Silicon Macs.

OAuth replaces in-app login

The default login flow now opens your browser for OAuth-based authentication instead of asking for username and password inside the app itself. This change directly addresses credential stuffing attacks, where bots try leaked username/password combinations against login forms. Password login is still available as a fallback, but OAuth is the default going forward.

Electron and build changes

The app upgraded from Electron 22 to Electron 25, which brings Chromium and Node.js updates under the hood. 32-bit builds are fully retired. This affects very few users at this point, but it's a clean break that simplifies the build pipeline.

On Linux, the installer now ships an AppImage instead of the previous format. There are also new arm64 Linux builds for the first time, which is relevant for developers on ARM-based Linux machines and single-board computers like Raspberry Pi.

Butler rebuilt with Go 1.24

Butler, the CLI tool that developers use to upload and manage game builds on itch.io, has been rebuilt with Go 1.24 and updated dependencies. This is a backend change that most developers won't notice directly, but it keeps the tool compatible with current systems and TLS requirements.

Why this matters

The itch.io desktop app is how a lot of indie game players discover and manage games outside of Steam. It handles downloads, automatic updates with delta patching, and offline play for browser-based games. The app had been relatively quiet on updates for some time, so this release signals renewed investment in the platform's client-side infrastructure.

The modernized build and release pipeline should also mean more frequent updates going forward. The team has already followed up with v26.9.0, which upgraded to Electron 40 and added sandboxing overhauls across all three operating systems.

References

• itch app v26.6.0 (+ butler updates)
• 2026 Update and Check-In
• itch.io app download

South Korea's Framework Act on the Development of Artificial Intelligence took effect on January 26, 2026. It's the world's first comprehensive AI law, and Article 36 has direct consequences for game developers everywhere. Any AI-generated content in a game sold in Korea, whether that's NPC dialogue, character art, or background music, must be labeled "AI-generated."

Overview of South Korea's AI Basic Act and its scope

What the law requires

Article 36 mandates a "duty to label generative AI" content. Game developers must clearly mark any content created with AI tools using watermarks or notices. This applies to illustrations, voice lines, story text, localization, background music, video sequences, and audio effects. Partial AI use counts too. If you used an AI tool to generate a texture that a human artist then refined, the law still expects disclosure.

The intent is to protect users' right to know what was made by a human and what wasn't. But game developers have raised concerns about breaking player immersion. Putting "AI-generated" labels on NPC dialogue in a fantasy RPG creates an odd experience.

It applies to foreign studios

The law has extraterritorial provisions. Global platforms like Steam and Epic Games must designate domestic Korean representatives and comply with the same requirements as Korean companies. If your game is available to Korean players, you're covered regardless of where your studio is located.

There's a one-year grace period before penalties kick in, so enforcement starts in early 2027. But the compliance work needs to start now for studios with Korean distribution.

"Permit first, regulate after"

One notable aspect: the law includes a "permit first, regulate after" principle for experimental AI applications. You don't need pre-approval to use AI in development. The government is using a task force approach, refining policies based on real-world deployment feedback rather than trying to anticipate every scenario upfront.

Steam narrows its AI disclosure

Separately, Valve updated Steam's AI disclosure policy in January 2026. The change draws a clearer line between what needs disclosure and what doesn't. "Player-consumed" content, meaning assets that ship in the game or are generated during gameplay, requires disclosure. Development tools like code assistants, debugging software, and AI-powered dev utilities are explicitly exempt.

Valve's reasoning: efficiency tools used during development don't change what appears on screen for the player. Around 8,000 Steam games disclosed AI use in the first half of 2025 alone, an 8x increase from roughly 1,000 for all of 2024.

The EU is next

The EU AI Act hits full enforcement on August 2, 2026. Penalties go up to 35 million euros or 7% of global annual revenue, whichever is higher. The EU requirements will apply across all 27 member states, creating an even larger compliance surface than Korea.

Between Korea's law, Steam's updated rules, and the approaching EU enforcement deadline, AI disclosure in games has gone from a niche compliance topic to an operational requirement for any studio with international distribution.

References

• Korea's AI Framework Act impact on games (Inven Global)
• Steam AI disclosure rules update (Everyday AI Blog)
• Valve tweaks AI disclosure rules (Game Developer)
• Korea AI Act video games labeling (The Korea Herald)

GDC 2026, rebranded this year as the "GDC Festival of Gaming," drew roughly 20,000 attendees. That's a 30% drop from the previous year and the lowest attendance since 2011. The State of the Industry survey, based on responses from over 2,300 game professionals, painted a picture of an industry in transition.

GDC 2026 recap: AI, user-generated content, and the new funding reality

Unreal passes Unity

For the first time, Unreal Engine is the primary engine used by survey respondents, at 42% compared to Unity's 30%. This is a notable shift. Unity held the top spot for years, but the fallout from the 2023 runtime fee controversy, ongoing layoffs, and a perception that Unity's tooling has stagnated all contributed to the change.

AI: companies adopt, developers push back

The AI numbers tell a split story. 52% of companies now use generative AI in some part of their workflow. But 52% of developers also say generative AI has negatively impacted the industry, up from 30% just a year ago. Only 36% of developers personally use AI tools, and just 7% see a net positive impact.

The tension is straightforward: studios see efficiency gains, while individual developers worry about job displacement and quality erosion.

Industry roundup covering layoffs and developer sentiment at GDC 2026

Layoffs continue

28% of respondents experienced a layoff in the past two years. Among US-based developers, that number rises to 33%. Two-thirds of AAA studio respondents said their employer conducted layoffs in the past 12 months.

The most troubling stat: 48% of laid-off developers haven't found new work in the industry. That's a lot of experienced talent leaving the field.

Unionization support grows

82% of US-based respondents support unionization. That's a strong and consistent signal. Whether it translates to organizing activity at specific studios remains to be seen, but the sentiment is no longer fringe.

What this means

The lower attendance isn't just about cost-cutting travel budgets. It reflects a broader contraction. Studios are smaller, teams are leaner, and hiring has slowed. The engine shift from Unity to Unreal suggests developers are voting with their feet on tooling quality and corporate trust.

The AI divide will likely widen before it narrows. Companies will keep adopting AI tools. Developers will keep pushing back, especially as long as the job market stays tight.

References

• GDC 2026 attendance drops 30% to 20,000 visitors
• GDC 2026 State of the Game Industry survey
• Game Developer Podcast: Analyzing the 2026 State of the Industry Report

NVIDIA unveiled DLSS 5 at GTC 2026, calling it a "GPT moment for graphics." The new technology uses generative AI to reconstruct lighting and materials in real-time, going far beyond the frame generation and upscaling of previous DLSS versions. CEO Jensen Huang described it as "neuro rendering," a fusion of traditional 3D graphics and AI that adds photorealistic detail to game frames.

The gaming community's response was overwhelmingly negative.

NVIDIA's DLSS 5 Zorah tech demo showing neural rendering at 4K

The "yassification" problem

Gamers quickly noticed that DLSS 5 was altering character faces. In demo footage of Resident Evil Requiem, the character Grace Ashcroft appeared noticeably different with DLSS 5 enabled. Her face looked smoother, more polished, and less natural. Critics described it as "plastic, airbrushed, and weirdly over-enhanced." Similar issues appeared in Hogwarts Legacy demos, where an older woman's face became uncanny and artificially smooth.

The term "yassifying" took off on social media, comparing DLSS 5's face smoothing to heavily filtered Instagram selfies. YouTube comments on NVIDIA's reveal were described as "almost 100% negative." Memes showing before-and-after comparisons spread rapidly.

The core complaint wasn't about performance. Previous DLSS versions were praised specifically because they were invisible. They made games run faster without visibly changing how they looked. DLSS 5 visibly changes artistic intent, and that's where the backlash concentrated.

Hallucinations in generated frames

Testing revealed that DLSS 5's generative approach could produce hallucinated details in frames. Because the system uses AI inference to reconstruct elements rather than simply upscaling existing pixels, it sometimes generates details that don't exist in the source material. This is the same fundamental limitation that affects all generative AI models, now showing up in real-time game rendering.

NVIDIA's shifting explanation

NVIDIA initially presented DLSS 5 as having "3D scene understanding," implying the AI model understood the geometry and materials of the scene. An NVIDIA engineer later clarified that DLSS 5 actually operates on 2D frame data combined with motion vectors. It doesn't have access to the full 3D scene graph. The correction was subtle but significant. It means DLSS 5 is making educated guesses about lighting and materials from flat images, not working with actual 3D information.

NVIDIA responds to the DLSS 5 controversy

Studio reactions

Larian Studios, the developer behind Baldur's Gate 3, reportedly pulled back from some generative rendering tools for their next project after fan backlash. When the studio behind one of the most acclaimed recent RPGs steps away from a technology, it sends a signal to the rest of the industry.

Jensen Huang pushed back on critics, saying they were "completely wrong" and that DLSS 5 preserves artistic control because developers can fine-tune the results. But the demos shown at GDC didn't support that claim. The visible face alterations in multiple games suggested the technology overrides artistic decisions by default.

Why it matters

DLSS 5 became GDC 2026's biggest flashpoint because it forced a question the industry hadn't fully confronted: where is the line between AI as a performance tool and AI as an aesthetic decision-maker? DLSS 1 through 4 stayed firmly on the performance side. DLSS 5 crossed into making visible creative choices about how characters and scenes should look. For many players and developers, that's a fundamentally different product.

Digital Foundry's team published an initial enthusiastic analysis, then released a follow-up acknowledging they "posted too soon" and should have waited for broader feedback before praising the technology.

References

• The Verge: DLSS 5 AI 3D rendering
• TechSpot: NVIDIA confirms DLSS 5 relies on 2D frame data
• Windows Central: DLSS 5 backlash
• NVIDIA GTC 2026 blog

Microsoft dropped a big announcement at GDC 2026: every game shipping on Xbox now gets free access to PlayFab's cross-platform backend services. No Azure subscription. No per-user fees. The new offering is called PlayFab Foundation Mode, and it covers unified player accounts, cross-platform saves, matchmaking, economy systems, and live service management across Xbox, PlayStation, Nintendo Switch, Steam, Epic, and mobile.

Xbox on how indie developers are driving gaming's biggest ideas

What Foundation Mode includes

PlayFab Foundation Mode covers seven service pillars. Identity gives players unified accounts across platforms. Progression handles cross-platform saves, profiles, and statistics. Community provides friends lists, guilds, leaderboards, and real-time chat. Multiplayer includes lobbies, matchmaking, and networking through Azure infrastructure. Live Service Management covers title data, news feeds, push notifications, and custom backend logic. Economy manages catalogs, virtual currencies, inventory, and bundles. Game Data Stream provides telemetry and player interaction analytics.

Previously, this was a paid service. Now any developer who commits to shipping on Xbox gets it at no cost. You don't even need to wait for launch to start building. Access is available as soon as you create a game in Partner Center and request it through PlayFab Game Manager.

The onboarding overhaul

Xbox also announced that developer onboarding has been reduced from 30 days to about 30 minutes. The new process is modular, meaning you can start building before full approval. Automated agreements cut handling time by over 90%. Public documentation no longer requires an NDA. Upload speeds improved 2x to 13x.

This is a direct response to the common complaint that getting a game onto Xbox was slow and bureaucratic compared to Steam or the Epic Games Store.

ID@Xbox and indie success

The ID@Xbox program got some attention too. Microsoft reported that indie games in the program generated "hundreds of millions of dollars" in sales on the Xbox Store over the past year. That's a real number attached to a program that's been running for 13 years.

Xbox also made changes to the Store's algorithmic discovery to prevent bundle flooding, a practice where bundles would dominate category listings and push individual games down. The new Indie Selects program provides curated editorial spotlight for indie titles.

What this means for small studios

For indie developers, the PlayFab announcement removes one of the biggest cost barriers to cross-platform multiplayer. Building your own backend for player accounts, matchmaking, and cross-platform saves is expensive and time-consuming. Having that provided for free, on infrastructure that scales through Azure, changes the math on whether cross-platform is worth the investment.

The catch is you need to ship on Xbox. But given that Game Pass provides a viable distribution channel for indie games, and the onboarding friction has been reduced, that's a lower bar than it used to be.

References

• ID@Xbox at GDC 2026
• Introducing Foundation Mode for PlayFab
• What's Changed in Xbox Development
• Pure Xbox coverage

Chrome 146 landed with a feature that significantly widens WebGPU's hardware reach: compatibility mode. Until now, WebGPU required a GPU that supports Vulkan, Metal, or Direct3D 12. Compatibility mode removes that requirement by running on OpenGL ES 3.1 and Direct3D 11 instead.

Chrome for Developers introduction to WebGPU, the API that compatibility mode now brings to older hardware

How it works

Developers opt in by requesting a compatibility-level adapter:

const adapter = await navigator.gpu.requestAdapter({
  featureLevel: "compatibility"
});

The returned adapter and device then run WebGPU through legacy graphics backends. Applications that already work with core WebGPU will work in compatibility mode without changes, though some advanced features have restrictions.

What this means in practice

A lot of laptops and mobile devices still ship with GPUs that only support OpenGL ES 3.1 or Direct3D 11. These devices were completely locked out of WebGPU before. Compatibility mode brings them in.

Google is starting with Android and exploring support on ChromeOS and Windows next. This matters for web games and 3D applications that need the widest possible device coverage.

Other WebGPU additions in Chrome 146

Transient attachments. The new TRANSIENT_ATTACHMENT texture usage flag tells the GPU that a render target only needs to exist during the render pass. On tile-based GPUs (common on mobile), this keeps data in fast tile memory and avoids expensive writes to main memory.

texture_and_sampler_let WGSL extension. This allows textures and samplers to be assigned to let variables in shader code. It's groundwork for bindless rendering patterns, where shaders can access textures without explicit bind group entries.

W3C Candidate Recommendation

Separately, the W3C published WebGPU as a Candidate Recommendation Draft on March 3, 2026. This is a standards milestone. The GPU for the Web Working Group expects to demonstrate implementation of each feature in at least two deployed browsers on top of modern GPU system APIs before moving to full Recommendation status.

WebGPU is now available in Chrome, Firefox (since version 141 on Windows), and Safari 26. With compatibility mode lowering the hardware floor and the W3C standardization advancing, WebGPU is on track to become the default graphics API for the web.

References

• What's New in WebGPU (Chrome 146)
• WebGPU W3C Candidate Recommendation Draft
• WebGPU Compatibility Mode Fundamentals

Meshy announced a new experimental division called Meshy Labs at GDC 2026, along with its first game: Black Box: Infinite Arsenal. It's a survivor-roguelike where the gameplay mechanics themselves are generated by AI at runtime. Not the assets. The actual weapon physics, elemental effects, trajectories, and damage parameters.

An overview of Meshy's AI-powered 3D model generation capabilities

How it works

Players type text prompts to describe the weapons they want. A "Designer Agent" takes those prompts and assembles atomic mechanics into playable weapons in real time. There's no fixed weapon database, no preset loadouts, and no hardcoded builds. Every run produces unique gameplay based on what the player asks for and how the AI interprets it.

The result is that two players can play the same level and have completely different experiences. One might describe an "ice beam that bounces off walls" and get exactly that. Another might type "gravity hammer that pulls enemies in" and get a different set of physics entirely.

What Meshy is framing

CEO Ethan Hu positioned this as a shift from "AI for 3D" (which is what Meshy is known for) to "AI for Fun." The idea is that AI shouldn't just speed up asset creation. It should become part of the gameplay loop itself, where players co-create the experience.

That's a big claim. Whether it works at scale depends on how well the Designer Agent handles edge cases, balances difficulty, and avoids generating broken or unplayable weapons. A demo was playable at GDC booth #941, and the game is listed on Steam as coming soon.

Business milestones

The numbers behind the announcement are significant. Meshy doubled its annual recurring revenue to $30M in just three months. The platform has passed 10 million users globally. The company also released Meshy-6, its latest generative model, which features improved geometry for both organic and hard-surface models and a new low-poly generation mode.

Meshy AI generating 3D models from text and images in seconds

Why it matters for game developers

Most AI-in-games conversations focus on asset pipelines: generating textures, 3D models, dialogue, or level layouts. Black Box: Infinite Arsenal is the first public example of AI generating the mechanics themselves. If the approach works, it could open a new design space where gameplay emerges from player intent rather than developer-authored systems.

It's early. But the fact that a company with $30M ARR and 10 million users is putting resources behind this makes it worth watching.

References

• Meshy unveils Meshy Labs at GDC 2026
• Meshy-6 launch post
• Black Box: Infinite Arsenal on Steam

Lightricks has released LTX 2.3, an open-source video generation model that produces native 4K video at up to 50 FPS with synchronized stereo audio in a single forward pass. It's the first open-source model to combine high-resolution video and audio generation together, and it runs on consumer hardware.

Official LTX 2.3 introduction from Lightricks

What's in the model

LTX 2.3 is a 22-billion-parameter DiT (Diffusion Transformer) that handles video and audio generation as a unified task. It supports text-to-video, image-to-video, audio-to-video, video-to-video, and depth conditioning. You give it a text prompt and get back a video clip with matching sound. No separate audio generation step, no post-syncing.

The model generates at 24, 25, 48, or 50 FPS in clips up to 20 seconds. It handles both horizontal and native portrait (9:16) aspect ratios, which matters for anyone targeting social media formats. Audio output is 24 kHz stereo.

What changed from LTX 2

The biggest improvements over the previous version are sharper details and better motion. LTX 2 had a tendency to produce somewhat static visuals. LTX 2.3 addresses that with a rebuilt VAE (variational autoencoder) for crisper textures and more natural movement.

Prompt understanding got a significant upgrade too. The text connector is now 4x larger with added gated attention, which means the model more actively references different parts of your prompt throughout the generation process. If you describe a specific camera movement while a character performs an action, it's better at maintaining both simultaneously.

Other improvements include cleaner audio with fewer artifacts, motion control features like first/last frame guidance, and camera effects including dolly, jib, and focus shift.

Hands-on look at what LTX 2.3 can produce

Performance

Lightricks claims the model runs 18x faster than competing models on H100 GPUs. On consumer-grade hardware, it can produce a 5-second clip at 24 FPS in about 4 seconds. That's fast enough for iterative work where you're generating multiple takes and picking the best one.

Why this matters for game development

Open weights mean you can run LTX 2.3 locally without API costs or usage limits. For game developers, that opens up several use cases: generating cutscene prototypes, creating trailer footage during pre-production, producing marketing materials, or rapid-prototyping cinematic sequences before committing to full production.

The combined video-plus-audio output is particularly useful for game trailers and cinematics where you'd otherwise need to sync audio separately. Being able to iterate on visual and audio tone simultaneously, in a single generation pass, compresses the feedback loop considerably.

The model is free to use under a 10M annual revenue threshold. It's available through the Lightricks API Playground, ComfyUI, PyTorch, and third-party platforms like Replicate.

References

• Lightricks: LTX 2.3 official page
• ClipRise: LTX 2.3 announcement
• Hugging Face: LTX 2.3 model weights
• Replicate: LTX 2.3 Pro

PlayCanvas Engine v2.17.0 landed with a focused set of improvements that push Gaussian Splatting forward on WebGPU. The release adds half-precision spherical harmonics, a GPU-driven sorting and culling pipeline, and a global splat budget for managing scene-wide LOD.

SuperSplat walkthrough with PlayCanvas CEO Will Eastcott on editing and optimizing Gaussian Splats

f16 spherical harmonics

The big technical addition is half-precision (f16) spherical harmonics for Gaussian Splatting when running on WebGPU. Spherical harmonics encode view-dependent color information for each splat, and they're one of the heaviest data structures in a splatting pipeline. Switching from 32-bit to 16-bit floats cuts memory usage roughly in half with minimal visual difference.

This only works on WebGPU because WebGL doesn't natively support f16 buffers. It's another example of WebGPU enabling optimizations that weren't practical before.

GPU-driven sorting and frustum culling

v2.17.0 introduces a GPU-driven sorting and frustum culling pipeline for splats on WebGPU. Previously, splat sorting happened on the CPU, which became a bottleneck with scenes containing millions of Gaussians. Moving this to compute shaders means the GPU handles both sorting and culling without round-tripping data back to the CPU.

The frustum culling step discards splats outside the camera view before they're sorted, reducing the workload further.

Global splat budget

A new global splat budget system provides scene-wide Gaussian Splat LOD management. When your scene contains multiple splat objects, you can now set a total splat count for the entire scene. The engine automatically distributes the budget across objects based on distance and screen coverage.

This is practical for applications that compose multiple scanned environments or objects into a single scene, like virtual showrooms or architectural walkthroughs.

Self-hosting Gaussian Splatting models with the PlayCanvas viewer

Color Enhance and rendering updates

The release adds a new Color Enhance post-effect with midtones control, plus half-precision bloom shaders for better performance. Material.meshInstances changed from an Array to a Set, which speeds up instance removal operations. Front face support was also added to the render pipeline.

v2.16.0 laid the groundwork

The prior release (v2.16.0, February 3) added 2D Gaussian Splatting (2DGS) support and WGSL subgroup support. 2DGS uses flat discs instead of 3D ellipsoids, which can produce sharper surface reconstructions for planar geometry like walls and floors. The WGSL subgroup support enables wave-level operations in compute shaders, allowing faster parallel reductions on supported GPUs.

Together, v2.16 and v2.17 represent a significant push toward making PlayCanvas the go-to engine for web-based Gaussian Splatting.

References

• PlayCanvas Engine v2.17.0 release
• PlayCanvas Engine v2.16.0 release
• PlayCanvas Gaussian Splatting documentation
• SuperSplat editor

Steam Next Fest wrapped its February 2026 edition with over 2,200 participating games and more than 3,500 playable demos. The data from this round tells a clear story about what works for indie discovery on Steam and what doesn't.

A look at standout indie games from Steam Next Fest February 2026

The tier system

Analysis by presskit.gg established clear performance benchmarks based on wishlists earned during the festival:

The median game earned around 460 wishlists during the festival. That's the realistic baseline.

What performed well

Co-op games overperformed. Content creators gravitated toward games they could play with friends on stream, which drove organic visibility. Social play translates into social media reach.

Survival and crafting titles stayed strong. No signs of audience fatigue in this genre despite how crowded it feels from the developer side.

What struggled

Narrative games and visual novels had a hard time. The festival format rewards quick, visually distinct experiences that photograph well in thumbnails. Story-driven games need more context to communicate their appeal.

Roguelikes faced saturation but could still break through with strong differentiation.

The momentum effect

The most striking finding: games that entered Next Fest with fewer than 1,000 existing wishlists rarely broke into higher tiers. The games that hit Diamond almost always started with 10,000+ wishlists already in place.

Next Fest amplifies existing momentum. It doesn't create it from nothing. The event rewards games that already have an audience. If you're launching cold, the festival alone probably won't change your trajectory.

What this means for developers

The data reinforces that discovery on Steam increasingly favors games with pre-existing visibility. Capsule art that reads at small sizes matters. Content creator relationships before the festival matter. Having a community before you show up matters.

For games that don't have that built-in momentum, alternative discovery paths become more important. We wrote about this dynamic in our Steam Next Fest strategy guide.

References

• presskit.gg analysis
• The Indie Informer top demos

Rémi Verschelde, Godot co-founder and one of the engine's core maintainers, has publicly described the surge of AI-generated pull requests hitting the project as "draining and demoralizing."

The problem

Maintainers now have to second-guess nearly every pull request from new contributors. The AI-generated submissions often look plausible at first glance but contain fundamental errors, broken logic, or code that doesn't make sense in context. The time spent identifying and rejecting these PRs takes away from reviewing legitimate contributions.

Godot currently has 4,681 open pull requests on GitHub. Distinguishing honest human mistakes from AI-generated nonsense has become, in Verschelde's words, extremely difficult.

Why it's happening

Godot's popularity surged after Unity's 2023 Runtime Fee controversy drove developers to explore alternatives. More users means more potential contributors, and a growing number of those contributors are using AI tools to generate code submissions without understanding what the code does.

Godot has contribution guidelines that require disclosure of AI assistance, but these are frequently ignored. The submissions aren't typically malicious. They come from people who genuinely want to contribute but are leaning on AI tools that produce confidently wrong code.

What Verschelde proposed

His primary recommendation was more funding to hire additional maintainers who can handle the review volume. He acknowledged that using AI to detect AI-generated PRs would be "horribly ironic" and expressed reluctance to automate the process.

Other possibilities floated include moving off GitHub or implementing community trust-rating systems, though both options risk raising barriers for genuine new contributors, which runs counter to the open-source ethos.

The broader pattern

This isn't only happening to Godot. Open-source projects across the software ecosystem are reporting similar patterns. AI tools make it trivially easy to generate a pull request. They don't make it easy to generate a useful one. The cost of submitting has dropped to near zero while the cost of reviewing remains high.

References

• Game Developer
• The Register
• PC Gamer

Nintendo held its Indie World Showcase on March 3, and it made one thing clear: the Switch 2 is launching with indie games front and center. The 15-minute presentation featured a mix of day-one shadow drops, ports of PlayStation hits, and new announcements from established indie studios.

Nintendo Indie World Showcase, March 3, 2026

Day-one launches

Three games dropped the same day as the showcase. Blue Prince is a genre-bending puzzle game set in a mansion where rooms shift and rearrange. It uses the Switch 2's Joy-Con mouse controls for point-and-click navigation. Rotwood is a co-op action brawler from Klei Entertainment, the team behind Don't Starve. It supports up to four players locally or online with deep combat combos and gear customization. Minishoot' Adventures is a twin-stick shooter mixed with classic adventure game exploration.

Shadow-dropping three titles on the same day as the announcement is an aggressive move. It signals that Nintendo coordinated closely with indie studios to ensure the Switch 2 had playable content from day one.

Blue Prince launch trailer for Nintendo Switch 2

Upcoming titles

The showcase also revealed release windows for several anticipated games. Heave Ho 2 from Devolver Digital brings the chaotic co-op party game online for the first time, coming summer 2026. Moonlighter 2: The Endless Vault continues the roguelite shopkeeping formula, slated for 2026. Denshattack! is a rhythm-action game arriving June 17. Mixtape and several other titles round out the lineup through fall 2026.

Kena arrives on Switch 2

One of the biggest announcements was Kena: Bridge of Spirits coming to Switch 2 this spring. Kena was a breakout PlayStation indie hit when it launched in 2021, known for its Pixar-quality visuals and challenging combat. The Switch 2 version includes the Anniversary DLC with Charmstones, Spirit Guide Trials, new outfits, and a New Game+ mode with redesigned combat encounters.

Getting Kena on a Nintendo platform is a statement about the Switch 2's hardware capabilities. The original game pushed the PS5, so a successful port demonstrates that the new hardware can handle visually demanding indie titles.

What it means for the platform

The original Switch built its library on indie support. Hollow Knight, Celeste, Stardew Valley, and dozens of other indie titles found massive audiences on the platform. Nintendo is clearly trying to replicate that dynamic from the start with the Switch 2.

For indie developers, the message is straightforward: Nintendo wants your games at launch, not six months later. The shadow drops and coordinated release timing suggest that the ID and certification process has been streamlined compared to the original Switch's early days.

References

• Nintendo Indie World Showcase (official)
• The Verge coverage
• Nintendo Life roundup

Roblox opened the beta for its 4D creation tool on February 4. The tool generates interactive 3D objects from natural language prompts, and unlike standard text-to-3D tools, these objects have real physics and player-driven behavior built in.

Roblox VP of AI explains the Cube foundation model and 4D creation

What "4D" means here

The "4D" in Roblox's naming refers to interactivity over time. A 3D model is a static mesh. A 4D object is a mesh with behavior. You prompt "a red sports car" and get a drivable vehicle with four independently spinning wheels, a body that responds to collisions, and controls that react to player input.

The system uses "schemas" that define object structure and behavior. At launch, two schemas are available:

• Car-5: Five-part vehicles with working wheels, steering, and collision physics
• Body-1: Single-piece objects like boxes, sculptures, or furniture with basic interactive properties

The numbers

During early access (starting November 2025), players generated over 160,000 objects using the 4D tools. Users who engaged with 4D generation showed a 64% average increase in play time compared to baseline.

A showcase experience called "Wish Master" by developer Laksh lets players generate and drive cars, fly planes, and control dragons, all created through prompts in real time.

A creator testing the 4D AI creation tools inside Roblox

What's coming next

Roblox plans to let creators build custom schemas for more object types beyond cars and simple bodies. They're also working on reference-image-to-3D generation and what they call "real-time dreaming," an AI world model approach that generates dynamic environments and objects on the fly.

This builds on their earlier Cube 3D model (released March 2025), which was open-sourced and generated over 1.8 million 3D objects on the platform.

References

• TechCrunch
• Reuters
• Roblox Cube announcement

Unity CEO Matthew Bromberg announced that Unity will demo a new AI beta at the GDC Festival of Gaming (March 9-13) that lets creators "prompt full casual games into existence with natural language only."

Unity CEO Matthew Bromberg discussing AI strategy and the platform's direction

What Unity said

Bromberg described AI-driven authoring as Unity's "second major area of focus for 2026." The tool will be native to the Unity platform, which means generated prototypes can be iterated into production-quality games without switching environments.

Under the hood, Unity says it combines the company's understanding of project context and runtime with frontier models from OpenAI (GPT) and Meta (Llama). The AI assistant will be web-accessible to reduce friction.

Bromberg estimated this could bring "tens of millions of more people creating interactive entertainment" through AI-enabled development tools.

What we know and don't know

Known: The beta exists and will be shown at GDC. It targets casual games. It uses natural language input. It's integrated with Unity's runtime and context system. Enhanced in-app purchase commerce tools are coming alongside it, with general availability in Q2 2026.

Unknown: What "full casual games" means in practice. Whether generated games are editable at the code level or only through continued prompting. How it handles asset generation versus code generation. What the quality ceiling looks like. How it compares to existing text-to-game tools.

Why it matters

Unity is the most widely used game engine in the world. When Unity puts AI-first game creation at the center of its product strategy, it signals that the market has moved past the experimental phase. This is the same company that faced developer backlash over the 2023 Runtime Fee. They're betting AI authoring is what wins developers back.

GDC runs March 9-13 in San Francisco. The 2026 event is being restructured as a "Festival of Gaming" with simplified passes starting at $649 for five days.

References

• The Verge
• Game Developer
• GDC Festival of Gaming 2026

Godot 4.6 landed on January 27, following through on the Q1 2026 preview we covered earlier. The release is one of the most feature-dense in the engine's history, and the changes matter for web developers specifically.

What shipped

Jolt Physics is now the default. Every new 3D project in Godot 4.6 uses Jolt, the same physics engine behind Death Stranding 2. Jolt has been available as an experimental option since 4.4, but this release makes it the standard. For web exports, Jolt's performance profile is noticeably better than the old GodotPhysics backend.

Modern editor theme. The new default theme, based on the community-favorite Godot Minimal Theme by passivestar, replaces the classic blue-tinted interface. Clean lines, improved spacing, grayscale palette for color-sensitive work. It's a visual refresh that keeps focus on the viewport.

Unified docking system. Panels can now be moved between sides and bottom freely, and most docks support floating windows. You can rearrange the editor layout to match how you actually work.

Screen Space Reflections overhaul. The SSR system got a full rewrite with improved roughness handling, better visual stability, and a half-resolution rendering mode for performance on constrained hardware.

LibGodot. You can now embed Godot as a library inside other applications. Initial support covers Linux, Windows, and macOS. This opens the door for using Godot as a rendering or game logic layer inside custom tools, launchers, or entirely different applications.

Other changes worth noting: unique node IDs that survive scene reorganization, delta encoding for smaller patch updates, rotatable TileMap tiles, in-game speed controls for testing, and C++ tracing profiler support.

What it means for web developers

Godot's web story has improved every release since 4.3 introduced single-threaded export. The 4.6 release continues that trajectory. Jolt's better performance matters when you're running physics in a browser. The SSR half-resolution mode helps on mobile web. And LibGodot's embedding capability could eventually let you run Godot scenes inside existing web applications, though web support for LibGodot hasn't been confirmed yet.

References

• Godot 4.6 release page
• Game Developer coverage

Cinevva is now live—a new way to discover indie games through short gameplay reels and instant browser play.

The problem we're solving

• Store pages don't show the fun. Screenshots and trailers are marketing, not gameplay.
• Too many clicks to play. Download, install, load, tutorial... friction kills interest.
• Great games stay buried. Indies can't compete with AAA marketing budgets.

How Cinevva works

Scroll. See. Play.

Watch 15-second gameplay clips. When something looks fun, launch it instantly in your browser. No downloads. No installs. No waiting.

Why now

The timing is right:

• WebGPU is now in all major browsers, enabling better graphics
• WebAssembly performance keeps improving
• AI-generated content has created a trust crisis—players want to see real gameplay
• Instant play removes the friction that kills discovery

For players

Discover your next favorite game in seconds, not hours.

For creators

Get your game in front of players who actually want to play it.

Try it

Head to app.cinevva.com to start discovering games.

Godot 4.6 is projected for release in Q1 2026, following the engine's regular release cadence.

What to expect

Based on Godot's release policy and community discussions:

• Continued improvements to web export stability
• Potential groundwork toward a WebGPU backend (not supported in any release yet)
• Further WASM optimizations building on 4.5's SIMD support
• Bug fixes for web-specific issues reported in 4.5

Godot's web momentum

Since Godot 4.3's single-threaded web export (August 2024), each release has improved the web story:

• 4.3: Single-threaded export for easier hosting
• 4.4: Stability improvements
• 4.5: WASM SIMD by default

References

• https://docs.godotengine.org/en/latest/about/release_policy.html

China's game regulator, the NPPA, issued licenses for 1,771 games in 2025: 1,676 domestic titles and 95 imported ones. That's up roughly a quarter from 2024's 1,416 and the highest annual total since 2018, the year a months-long approval freeze rattled the entire industry and wiped out billions in market value. The single-month peak in 2025 reached 184 titles, with August alone clearing 173.

Why the count is the story

In China, a game legally cannot monetize without a license number, the banhao. When approvals slow or stop, as they did in 2018 and again during a 2021-2022 tightening, the whole domestic industry seizes up regardless of how good anyone's games are. The license pipeline is the master valve, and for years its uncertainty was the single biggest risk factor hanging over the sector.

2025 is the clearest signal in a long time that the valve is open and steady. Approvals didn't just rise in volume, they normalized in rhythm: a predictable monthly cadence instead of unpredictable droughts. Imports matter here too. Foreign titles had slowed to a trickle, and monthly issuance settling back to roughly nine or ten imported approvals a month means the door that had been mostly shut is open again. January 2025 alone cleared high-profile titles like Honor of Kings World and Star Resonance.

What it means for builders

A predictable regulator changes what studios can plan. When you can reasonably expect a license on a normal timeline, you can greenlight, budget, and staff a project without betting the company on a bureaucratic coin flip. Stability at the top of the funnel is what lets everything below it run. The record approval count and, more importantly, the steady cadence behind it, tell you the Chinese domestic market is planning aggressively again after years of defensive crouch.

For the rest of the world, this is the supply side of everything else in this timeline. Black Myth: Wukong, Marvel Rivals, the AI-tooling arms race between Tencent, NetEase, and miHoYo, all of it runs on a domestic market that's confident enough to invest. A healthy approval pipeline is the quiet precondition for the loud releases.

For Cinevva, the regulatory specifics are China's to navigate, but the pattern is universal: creation flourishes where the path to publish is predictable and open. Our version of that valve is a browser link. There's no gatekeeper deciding whether a creator's game gets to exist, no queue, no number to wait on. Different mechanism, same principle. The health of any creative ecosystem tracks how reliably a finished thing can reach an audience, and 2025 was the year China's answer to that question finally stabilized.

References

• Niko Partners: China's NPPA approved 1,771 video games in 2025
• Respawn: China approves 1,771 games in 2025, highest since 2018
• TechNode: China approves 173 game licenses in August, highest monthly total of 2025
• Game World Observer: China approves over 130 games in January, including Honor of Kings World

As of early 2026, WebGPU is supported by default in Chrome, Firefox, Safari, and Edge across desktop and mobile platforms. This marks a fundamental shift in what's possible for browser-based games.

What changed

• All major browsers now ship WebGPU by default
• Compute shaders enable GPU-accelerated physics, AI, and effects
• Modern rendering patterns that were impossible with WebGL are now viable
• Engine support from Three.js, Babylon.js, Unity, PlayCanvas, and others

What this means for developers

The conversation has shifted from "will WebGPU be supported?" to "how do I best use WebGPU?"

Key considerations:

• Keep WebGL fallbacks for older browsers and devices
• Test across browsers—implementations vary slightly
• Profile your specific scenes, not just demos
• Consider compute shaders for physics and particle systems

The ecosystem in 2026

• Babylon.js 8.0: Native WGSL shaders, no GLSL conversion needed
• Three.js r170+: Zero-config WebGPU imports
• Unity 6.1: WebGPU as a supported backend
• Godot: WebGPU backend in development

Looking ahead

Web games can now achieve visual quality that was previously reserved for native applications. The gap between "web game" and "real game" continues to narrow.

Tencent launched its Hunyuan 3D engine globally on November 26, 2025. Feed it a text description, an image, or a rough sketch and it returns a high-quality 3D asset in minutes, work that traditionally took days or weeks. There's a free tier of 20 generations a day for standard users, 200 starter credits for enterprises on the Hunyuan 3D Model API, and Tencent says more than 150 companies in mainland China had already wired the API into their pipelines before the global launch.

From research demo to infrastructure

Text-to-3D has been a research party trick for a couple of years: impressive in a paper, unusable in a real production because the meshes were messy, the topology was garbage, and nothing was game-ready. Hunyuan 3D is Tencent's argument that this phase is over. The lineup spans object generation for props and characters, a PolyGen model aimed at art-grade topology, and Hunyuan3D World models for building large interactive environments you can actually move through.

The free tier is the tell. You don't put a generous free tier on a research demo. You put it on a funnel. Tencent is treating generative 3D as infrastructure it wants sitting under as many pipelines as possible, gaming and beyond, into commerce, film effects, advertising, and social content. That's a platform play, not an experiment.

What it means for builders

Procedural geometry and hand-modeling were the two options for filling a 3D world, and both have a wall. Procedural gets you infinite boxes and cylinders that all read as placeholder. Hand-modeling gets you beautiful assets at the speed of one skilled human. A generative asset engine that outputs usable meshes in minutes is a genuine third path, and Tencent putting it behind a free tier means it's not a luxury reserved for studios with a Tencent contract.

There's a competitive read here too. Tencent, NetEase, and miHoYo are in what one analysis called a covert technological battle over the AI games era, and Hunyuan is Tencent staking the asset-generation layer. When the largest game company on the planet decides the 3D content pipeline should be a commodity utility, the pipeline becomes a commodity utility. Everyone downstream gets to build as if art assets are cheap, because they're about to be.

For Cinevva this is squarely the world we build in. Our creators already reach for a library of thousands of CC0 models so their games escape procedural-cube territory, and generative asset engines like Hunyuan are the next layer of the same idea: the look of a game should not be gated by whether you can model or afford someone who can. Tencent just made that case at global scale, for free, to 150-plus companies who took it live before the rest of the world even saw the announcement.

References

• Game Developer: Tencent launches AI-powered 3D asset tool to "empower creators"
• Tencent: Global launch of Hunyuan 3D engine
• PocketGamer.biz: Tencent rolls out Hunyuan 3D engine for AI-driven asset creation

WebGPU has moved from "future tech" to something you can realistically target for most players.

Why this matters for web games

• Better GPU access than WebGL for modern rendering patterns
• Compute shaders for effects, simulation, and GPU-side workloads
• A clearer path to desktop-quality visuals in the browser

Browser support timeline

• Chrome 113+ (May 2023)
• Firefox 141+ on Windows (2025)
• Safari 26 (September 2025)
• Edge follows Chrome

What to do as a developer

• Add capability detection and keep a WebGL fallback for older browsers
• Expect different performance profiles vs WebGL: measure on your real scenes
• Treat WebGPU as an opportunity to simplify pipelines

References

• https://web.dev/blog/webgpu-supported-major-browsers
• https://developer.chrome.com/blog/new-in-webgpu-113

Cinevva has been selected as one of the stars of this generation's World's Top Performing Startups in the digital science category by Shanghai's local investors ecosystem.

What this means

The recognition highlights Cinevva's mission to make game creation, discovery, and self-publishing accessible to everyone—across all channels and devices.

Our approach to AI

While the AI race intensifies, we're taking a different path: orchestrating state-of-the-art AI tools to serve our users' end goals—not replacing creators, but empowering them.

• Bring ideas to life in minutes
• Publish more, faster
• Express yourself through games
• Prompt from your phone or PC

Build a voice. Build an audience. Thrive.

Looking ahead

We're excited to explore partnerships with organizations like Tencent, Tencent Games universities, and other supporters in the space as we continue growing globally.

Source: Mariana's LinkedIn announcement

Godot 4.5 was released on September 18, 2025, with WebAssembly SIMD enabled by default for web exports.

Key web export improvements

• WASM SIMD by default: Single Instruction Multiple Data operations improve performance without manual configuration
• Single-threaded export remains default: Avoids Cross-Origin Isolation header requirements
• PWA support improved: Service worker fallback for COOP/COEP headers

PWA features

• High-resolution icons
• Display orientation settings
• Offline support via service worker caching
• Custom offline page option

Continued limitations

• C# not supported for web export (use GDScript or Godot 3.x for C#)
• Audio restrictions in browsers (user interaction required)
• Full-screen must be triggered by user input

References

• https://godotengine.org/releases/4.5/
• https://docs.godotengine.org/en/4.5/tutorials/export/exporting_for_web.html

Safari 26 marks Apple's full commitment to WebGPU, enabling it by default across all Apple platforms.

What's included

• macOS Tahoe (macOS 26)
• iOS 26 and iPadOS 26
• visionOS 26

Why it matters

Safari was the last major browser to ship WebGPU support. With this release, web games can target WebGPU across essentially all modern browsers without feature detection fallbacks for most users.

For web game developers

• Test your WebGPU games on Safari/WebKit
• Apple's Metal backend may have different performance characteristics than D3D12/Vulkan
• Mobile Safari now supports compute shaders

References

• https://webkit.org/blog/14879/webgpu-now-available-for-testing-in-safari-technology-preview/

Three.js r170 continues the WebGPU maturation with improved organization and broader format support.

Key changes

• WebGPU modules moved to addons for cleaner imports
• GLTFLoader improvements under WebGPU renderer
• Better compatibility with existing Three.js codebases

Import changes

import { WebGPURenderer } from 'three/webgpu';

For web game developers

The WebGPU path in Three.js is becoming more stable and production-ready with each release.

References

• https://github.com/mrdoob/three.js/releases

Firefox 145 extends WebGPU support to Apple Silicon Macs, an important milestone for cross-browser compatibility.

Support timeline

• Firefox 141: Windows support
• Firefox 145: macOS ARM64 support
• Linux and Android: still rolling out

Why it matters

With this release, Firefox joins Chrome and Safari in supporting WebGPU on the most common desktop configurations. Web game developers can now target WebGPU with higher confidence.

References

• https://www.mozilla.org/en-US/firefox/releases/

Unity 6.1, released in April 2025, adds WebGPU as a supported graphics backend for web builds.

What's included

• WebGPU backend for browser builds and Instant Games
• Improved web performance and stability
• Enhanced debugging tools for web exports
• Instant Games platform support expansion

Why it matters

Unity catching up on WebGPU means the most-used game engine now offers a modern path to browser deployment. Developers can target WebGPU for better performance while falling back to WebGL for compatibility.

Considerations

• Test early for build size and load times
• WebGPU support varies by browser
• Keep shaders and assets optimized
• Some browser APIs may not be implemented

References

• https://unity.com/releases/unity-6

GeoGuessr, the geography guessing game that built its audience as a browser game, launched on Steam in Early Access.

Why it matters

This demonstrates that successful browser games can expand to traditional gaming platforms. The path from web to Steam is viable for games that build an audience online first.

The browser-to-Steam model

• Build and iterate quickly on the web
• Grow an audience without install friction
• Expand to Steam for discoverability and features

References

• https://store.steampowered.com/app/2959080/GeoGuessr/

Babylon.js 8.0 was released on March 27, 2025, with a major architectural improvement: all core engine shaders are now available in WGSL.

Key features

• WGSL-first shaders: No GLSL-to-WGSL conversion needed for WebGPU builds
• Smaller bundle sizes: Eliminating the translation layer reduces download size
• IBL Shadows: Image-based lighting shadow support
• Area Lights: More realistic lighting options
• Node Render Graph (alpha): Visual render pipeline editing

Why WGSL matters

Previously, Babylon.js would convert GLSL shaders to WGSL at build time for WebGPU. With native WGSL support:

• Faster builds
• Smaller bundles
• More predictable behavior
• Better debugging

For web game developers

If you're using Babylon.js for WebGPU games, 8.0 is a significant upgrade. The reduced overhead makes WebGPU an even more attractive target.

References

• https://blogs.windows.com/windowsdeveloper/2025/03/27/announcing-babylon-js-8-0/
• https://forum.babylonjs.com

Cinevva won second place ($10,000) in the Dell for Startups pitch competition at SXSW 2025 in Austin.

The competition

Founder Institute Texas alumni pitched live on the Dell Experience stage at SXSW, competing for $25,000 in total prize money provided by Dell for Startups.

The pitch

Cinevva founder Mariana Muntean presented the platform's vision: making game creation and discovery accessible to non-technical creatives through browser-based tools and AI assistance.

Results

• 1st place ($15,000): Intuily (Austin)
• 2nd place ($10,000): Cinevva (Houston)

What this means

The recognition validates Cinevva's approach to democratizing game creation and discovery. The prize money supports continued development of the platform.

References

• Founder Institute announcement

Godot 4.4 was officially released on March 3, 2025, continuing the engine's investment in web as a first-class platform.

Web export improvements

• Continued refinements to single-threaded export (introduced in 4.3)
• Better WebGL 2.0 compatibility
• Improved audio handling in browsers

Known limitations

• C#/.NET projects still not supported for web export
• Some audio features (AudioStreamInteractive) have issues in web builds

References

• https://godotengine.org/releases/4.4/
• https://docs.godotengine.org/en/4.4/tutorials/export/exporting_for_web.html

DeepSeek released R1 on January 20, 2025, and the shock wasn't the benchmark scores. It was the price. R1 matched frontier reasoning models at around $0.07 per million input tokens, roughly 27 times cheaper than OpenAI's comparable model, and DeepSeek said it trained the thing for about $5.6 million on 2,000 GPUs in 55 days. NVIDIA lost $589 billion of market cap in a single day as the market repriced what compute was worth.

The part that matters for games

Most of the coverage was about markets and geopolitics. The part that matters for anyone building games is what happened next inside the Chinese studios. Within weeks, NetEase CEO Ding Lei said DeepSeek-class models had lifted the company's game R&D efficiency by about 30 percent: mission-script generation that used to take two weeks was landing in three days, and associated labor costs dropped roughly 40 percent.

Read that again. This isn't a lab result or a demo. It's the head of one of the largest game publishers on earth putting concrete numbers on how much cheaper and faster production got, almost immediately, because the cost of a capable model fell off a cliff.

What it means for builders

For two years the AI conversation in games was about capability: whose model writes better dialogue, whose generates cleaner code, whose makes usable art. That race mostly resolved into a tie at the top. R1 moved the frontier to a different axis entirely, the same one the AI video world would spend all of 2026 discovering: cost.

When intelligence gets 20-plus times cheaper, the economics of every AI feature in a game pipeline invert. Things that were too expensive to run on every asset, every quest line, every player interaction become things you run by default. The question stops being "can we afford to put a model in this loop" and becomes "why isn't there one in every loop." A big publisher captures that as a 30 percent efficiency gain across a thousand-person org. A solo creator captures it as the difference between an idea being buildable and not.

That second version is Cinevva's whole thesis. We orchestrate frontier models so a single creator can build a game that would have needed a team, and every time the underlying cost of intelligence drops, the floor of what one person can make rises with it. DeepSeek didn't just embarrass a few US labs. It moved the price of the raw material that solo game creation is built from, and the people best positioned to spend that windfall aren't the ones with the biggest budgets. They're the ones who never had a budget at all.

References

• DeepSeek: R1 release notes
• Altman Solon: Three takeaways from the DeepSeek R1 release
• 36Kr: Who will rule the AI games era? Tencent, NetEase, and miHoYo in a covert technological battle

As of December 11, 2024, WebAssembly garbage collection (WasmGC) is available by default in all major browsers.

Browser support

• Chrome 119+: enabled by default
• Firefox 120+: enabled by default
• Safari 18.2: added support

Why it matters for web games

WasmGC makes it easier for languages with garbage collection (Kotlin, Dart, Java, etc.) to compile to WebAssembly without bundling their own GC. This means:

• Smaller WASM bundles
• Better performance for GC-heavy languages
• More options for game development toolchains

The practical impact

Games and engines written in languages that need GC can now target the browser more efficiently. Expect to see more diverse toolchains targeting WASM.

References

• https://web.dev/blog/wasmgc-wasm-tail-call-optimizations-baseline
• https://v8.dev/blog/wasm-gc-porting

NetEase Games launched Marvel Rivals on December 6, 2024, on PC, PlayStation 5, and Xbox Series X/S. It hit 10 million players in 72 hours, 20 million within ten days, and eventually peaked around 40 million. A team-based hero shooter, built in close collaboration with Marvel Games, became the breakout multiplayer launch of the year, and it came from a Chinese publisher, not from Blizzard or Valve.

Marvel Rivals "Rivals 'Til the End" official launch trailer

The hit

Overwatch defined the modern hero shooter and then spent years fumbling the goodwill. Marvel Rivals walked into that opening with a licensed roster people already loved, destructible environments, and a fast, generous content cadence. NetEase read the room precisely: the audience was there, the incumbent had drifted, and Marvel characters are the closest thing gaming has to a universal language. Forty million players in a few weeks is the result of getting all of that right at once.

The twist

Here's the part that stuck with the industry. Two months after this launch, NetEase laid off the US-based design team that helped build Marvel Rivals, even as the game was thriving. A record-setting hit was not enough to protect the people who made it. The development was consolidated back toward NetEase's China studios while the game kept climbing.

That's not a footnote. It's the clearest signal of the year about where the leverage in this business actually sits. The value a live-service hit throws off does not automatically flow to the team that shipped it, and being on the winning game is not the same as being safe.

What it means for builders

The hit and the layoff are one lesson, told from two ends.

From the top: a Chinese publisher out-executed the Western incumbents at their own genre, on their own platforms, in their own biggest media franchise. The idea that hero shooters or live-service polish are a Western competency is done. The talent, the tools, and the operational discipline are global now, and the next Overwatch-scale hit is as likely to come out of Guangzhou as out of Irvine.

From the bottom: even a 40 million-player success didn't buy the individual developers job security, because they were one input into a machine somebody else owned. That's the structural problem underneath the whole industry's layoff cycle. The people who make the thing rarely own the thing.

For Cinevva, that second half is the reason the platform exists the way it does. When a creator builds and publishes on Cinevva, the game is theirs, the audience is theirs, and the upside is theirs. We're betting the durable version of this industry is one where more of the value lands with the person who made the game instead of evaporating up a corporate ladder the moment the quarter closes. Marvel Rivals is proof that Chinese studios can win the biggest genres outright. It's also proof that winning the game and keeping your job are two different things.

References

• Neowin: Marvel Rivals crosses 20 million players
• Game Developer: Marvel Rivals topped 40 million players before US developers were laid off
• Gameranx: Marvel Rivals celebrates reaching 20 million players

Three.js r171 makes it easier than ever to use WebGPU with zero-config imports.

What changed

You can now import WebGPURenderer without special bundler configuration:

import { WebGPURenderer } from 'three/webgpu';

No polyfills, no extra setup, no bundler tweaks required.

Why it matters

Lower friction means more developers will try WebGPU. The Three.js team has been steadily reducing barriers to WebGPU adoption.

References

• https://github.com/mrdoob/three.js/releases/tag/r171

Epic Games launched Fab on October 22, 2024, consolidating multiple asset marketplaces into one unified platform.

What's included

• Unreal Engine Marketplace content
• Sketchfab Store assets
• Quixel Megascans (free through end of 2024)
• ArtStation Marketplace (migration planned for 2025)

Key details

• Revenue share: 88% to creators, 12% to Epic
• Format support: Unreal, Unity, Blender, Maya, Cinema 4D
• Planned support: Roblox and Minecraft formats coming

For web game developers

Fab assets work across multiple engines, including those that export to web. The unified search and licensing makes it easier to find and use assets legally.

References

• https://www.fab.com
• https://www.unrealengine.com/en-US/blog/fab-epics-new-unified-content-marketplace-launches-today

Short answer: yes, the Unity Runtime Fee is gone. Unity officially cancelled it on September 12, 2024, effective immediately, and it has not returned since. There are no per-install fees on any Unity game, on any version of the engine. If you came here to confirm the rumor, that's it: the fee is dead, and Unity went back to a normal seat-based subscription.

If you want the full story, the official source, and what Unity actually costs now, read on.

The official source

Unity announced the cancellation on its own blog, which is the authoritative confirmation developers keep searching for:

• Official announcement: unity.com/blog/news/pricing-and-packaging-updates

Everything below summarizes that announcement and the pricing that took effect afterward.

What the Runtime Fee was

In September 2023, Unity announced a "Runtime Fee" that would charge developers a small amount every time their game was installed, once it passed certain revenue and install thresholds. The reaction was immediate and overwhelmingly negative. Developers worried about install bombing, unpredictable bills, and being charged for installs they couldn't control. Several studios publicly threatened to leave Unity, and trust in the company's pricing took a serious hit.

What changed on September 12, 2024

Unity reversed course completely:

• No per-install fees for any game, on any version of Unity, for any customer
• A return to the seat-based subscription model developers already understood
• The free Personal tier revenue and funding cap was raised to $200,000 per year
• Unity 6 users can optionally remove the "Made with Unity" splash screen

In plain terms, Unity walked back the entire Runtime Fee and made the free tier more generous than before.

What Unity costs now

After the cancellation, Unity moved back to per-seat subscriptions. The pricing that took effect on January 1, 2025 and remains the structure today:

Pro rose roughly 8% and Enterprise roughly 25% when the new pricing landed. Always check the official pricing page for the current numbers before you budget, since subscription prices change over time.

Why developers still search for this in 2026

The Runtime Fee did lasting damage to how developers think about engine pricing. Even with the fee gone, many studios now plan around the question "what happens if my engine changes its terms again?" That's a fair thing to ask. The safest hedge is an engine whose license can't be revoked or repriced out from under you, which is why open-source engines saw a surge of interest after 2023.

Frequently asked questions

Is the Unity Runtime Fee still a thing?

No. Unity cancelled the Runtime Fee on September 12, 2024, effective immediately, and it has not been reinstated. There are no per-install charges on any Unity game today.

Did Unity ever actually charge the Runtime Fee?

No. The fee was announced in September 2023 to start in January 2024, but Unity delayed it amid the backlash and then cancelled it outright in September 2024 before it was ever applied.

Is Unity free in 2026?

Yes, for individuals and small studios. The Unity Personal tier is free as long as your revenue and funding stay under $200,000 per year. Above that, you move to a paid per-seat plan.

What replaced the Runtime Fee?

Nothing new. Unity simply went back to its previous seat-based subscription model, with a higher free-tier cap and slightly higher Pro and Enterprise prices.

Should I still trust Unity's pricing?

That's a judgment call. The fee is genuinely gone, and the current model is the conventional per-seat subscription Unity used for years. If pricing stability is a top concern, open-source engines like Godot, or web-first open-source options, remove the risk of a license change entirely.

If pricing predictability matters to you

The whole Runtime Fee episode was about one thing: developers losing control over the cost of shipping their own games. If that's what worries you, an open-source engine sidesteps it. Cinevva Engine is open-source and free, runs in the browser, and ships to mobile, desktop, Steam, and Discord from one codebase. No seats, no install fees, no terms that can change after you've built your game.

If you're actively weighing a switch, see our guide to Unity alternatives for web games, which picks an engine based on why you're leaving. For the broader field, see our comparison of the best web game engines for 2026, including how Unity WebGL stacks up against Godot, PlayCanvas, and the lightweight JavaScript frameworks.

Game Science launched Black Myth: Wukong on August 20, 2024, and the numbers didn't read like a first game from a studio nobody outside China had heard of. It peaked at 2,223,179 concurrent players on Steam, overtaking Cyberpunk 2077 as the most-played single-player launch in Steam history and settling in as one of the most-played Steam games of all time. It sold 10 million copies in three days, 18 million in two weeks, 20 million in a month, and crossed 25 million by January 2025.

Black Myth: Wukong official launch trailer

Why this one mattered

For a decade the story of Chinese games was mobile and free-to-play: gacha, live-service, monetization craft. Black Myth: Wukong is the opposite of all of that. It's a single-player, premium, buy-it-once action RPG built on Unreal Engine 5, rooted in the sixteenth-century novel Journey to the West, and it went straight to the top of the global charts on its own merits.

That's the shift. It isn't that a Chinese studio made money, they'd been doing that at enormous scale for years. It's that a Chinese studio made a big-budget, story-driven, critically embraced console-and-PC game and the whole world played it. The ceiling everyone assumed existed, the one where China does mobile and the West does AAA single-player, turned out not to be there.

What it means for builders

Two things follow, and they pull in opposite directions.

The first is inspiration. Black Myth started as a small team with an obsession and a mythology nobody else was mining. It didn't win by out-spending Sony or out-marketing Ubisoft. It won by making something specific and gorgeous from a culture that had never been rendered at that budget. Specificity travels. The most exportable thing a game can have is a point of view.

The second is scale, and it's a warning. A 2.2 million concurrent-player launch on a hand-crafted, years-in-the-making Unreal 5 production is not a template most people can follow. That path costs a fortune and takes half a decade, and for every Wukong there are studios that bet the same way and never shipped.

For Cinevva, the takeaway is to separate the ambition from the method. The ambition, a game with a real point of view that anyone in the world can pick up, is exactly right. The method, a multi-year AAA pipeline, is the part we're trying to make unnecessary. The reason to lower the cost and time of building a good-looking, good-feeling game is so that the next specific idea, the one that would never survive a five-year budget, gets made and gets found. Wukong proved the appetite is global. The open question is how many more of these exist that never get built because the only known route to them is a hundred-person studio and a decade.

References

• Forbes: Black Myth: Wukong is still putting up wild numbers a month after release
• Statista: Black Myth: Wukong global unit sales
• Game Rant: Black Myth: Wukong hits yet another sales milestone
• SteamDB: Black Myth: Wukong charts

Godot 4.3, released on August 15, 2024, brings back single-threaded web exports as a major quality-of-life improvement.

Why it matters

Multi-threaded web exports require special HTTP headers (COOP/COEP) that many hosting platforms don't support or make difficult to configure. Single-threaded exports work everywhere:

• itch.io and other game portals
• GitHub Pages and static hosts
• macOS and iOS Safari (which had issues with multi-threaded builds)

What changed

• Single-threaded export is now available and enabled by default
• Multi-threaded export still available when needed
• Reduced hosting friction for indie developers

References

• https://godotengine.org/releases/4.3/
• https://godotengine.org/article/progress-report-web-export-in-4-3/

Three.js r167, released in July 2024, enables WebGPU by default when the browser supports it.

What this means

When WebGPU is available, Three.js will automatically use WebGPURenderer instead of requiring manual setup. WebGL remains the fallback for browsers without WebGPU support.

Benefits

• Automatic detection: no code changes needed for WebGPU benefits
• Graceful fallback: WebGL still works everywhere
• Future-proof: apps automatically improve as browser support expands

For developers

Test your Three.js apps in WebGPU-capable browsers to ensure they work correctly with both renderers.

References

• https://github.com/mrdoob/three.js/releases/tag/r167

HoYoverse launched Zenless Zone Zero on July 4, 2024, simultaneously on PC, PlayStation 5, Android, and iOS. Three days later it had crossed 50 million downloads and hit number one in downloads across 27 markets, including China, the United States, Japan, South Korea, and Brazil. On mobile alone it earned $24.6 million in its first five days and close to $100 million by the end of July. About three-quarters of that mobile audience came from China and Japan.

Zenless Zone Zero official launch teaser

The polish is the product

ZZZ is a stylish urban-fantasy action RPG, and by HoYoverse's own standards its revenue was modest. It out-earned Genshin Impact and Honkai: Star Rail combined in that launch window, but it didn't touch the $55 million Star Rail pulled in its own first five days. Inside HoYoverse this counts as a soft opening.

That framing is the interesting part. A game that grosses $100 million in a month is a "weaker recent release" for this studio, because the bar it set with Genshin and Star Rail is that high. The bar isn't the art or the combat in isolation. It's the whole package: day-one launch on four platforms at once, a UI that feels the same on a phone and a PC, cross-save that just works, and a content cadence players can set a calendar by.

What it means for builders

The Chinese live-service studios have made cross-platform, launch-everywhere-at-once the default expectation, not the stretch goal. A game that ships on PC first and "comes to mobile later" now reads as unfinished to a huge slice of the global audience. Simultaneous is the new normal, and it's a brutal engineering bar.

For Cinevva this points at the exact seam we build on. Our games run in the browser, which means they're already everywhere a link opens, phone or desktop, no separate build to port and no store review to wait on. The lesson from ZZZ isn't "spend HoYoverse money." It's that reach and consistency across devices is what players now grade you on, and getting there shouldn't require a AAA budget and a four-platform release pipeline. That's the gap a web-native platform closes.

The other lesson is quieter. HoYoverse can call a $100 million launch a disappointment because its retention machine is so good that anything short of a phenomenon looks flat. The polish that earns that retention, the feel of every menu and every hit, is learnable and copyable. It's the part most small teams skip, and it's the part that compounds.

References

• Sensor Tower: Zenless Zone Zero slashes across download and revenue ranks
• VentureBeat: Zenless Zone Zero passes 50M global downloads in 72 hours
• PCGamesN: It's made more than $100 million, but ZZZ is still HoYoverse's weakest recent release
• Sportskeeda: Zenless Zone Zero reportedly made $96 million in July 2024

PlayCanvas announced integration with Sketchfab, allowing developers to browse, preview, and import 3D models directly within the PlayCanvas Editor.

How it works

• Browse Sketchfab's library from the Asset Store panel
• Preview models before importing
• License and author information preserved

Why it matters

For web game developers using PlayCanvas, this reduces friction when finding and adding 3D assets. No need to download, convert, and re-upload models manually.

Note

This is an integration, not an acquisition. Sketchfab remains owned by Epic Games and is now part of the Fab marketplace.

References

• https://blog.playcanvas.com/playcanvas-adds-sketchfab-integration

Unreal Engine 5.4 was officially released on April 23, 2024, with continued improvements to Pixel Streaming for web deployment.

Pixel Streaming 2 (experimental)

• New Pixel Streaming Player component
• Support for H.264, VPX, and AV1 encoders
• Display active pixel streams as media sources within UE projects

Why Pixel Streaming matters

Since Unreal dropped native HTML5 export, Pixel Streaming is the primary way to run high-fidelity UE content in browsers. The game runs on a server and streams video to the browser.

Trade-offs

• Pro: Full Unreal fidelity in browser
• Con: Requires server infrastructure and bandwidth
• Con: Latency-sensitive games may struggle

References

• https://www.unrealengine.com/en-US/blog/unreal-engine-5-4-is-now-available
• https://dev.epicgames.com/documentation/en-us/unreal-engine/pixel-streaming-in-unreal-engine

Babylon.js 7.0 was officially released on March 28, 2024, with significant performance improvements and new features.

Key highlights

• Performance gains across rendering pipeline
• WebGPU improvements including Snapshot Rendering (render bundles)
• Gaussian Splatting support for novel view synthesis
• Procedural geometry enhancements

Snapshot Rendering

Babylon.js's Snapshot Rendering uses WebGPU render bundles to cache and reuse GPU command lists, potentially achieving 10x CPU-side speedups for static geometry.

Why it matters

Babylon.js continues to push what's possible in browser-based 3D. For web game developers, 7.0 offers meaningful performance improvements without code changes.

References

• https://forum.babylonjs.com/t/introducing-babylon-js-7-0/49132
• https://blogs.windows.com/windowsdeveloper/2024/03/28/part-1-announcing-babylon-js-7-0/