Appearance
Play two games. Discover the models.
Compare AI-built games by how they feel to play, what they cost, and how reliably they work.
Play both. Pick your favorite.
Play freely. Sign in to record a vote without losing this matchup.
1Play A2Play B3Vote & reveal
Loading game…
Loading game…
Player preferences
These are early preferences from assigned matchups, with model names hidden until voting. A working game is not always the most enjoyable one. Technical results are measured separately below.
Early evidence · 4 votes · 3 voters · 8 builds. More independent players and builds are needed before naming a winner.
| Model | Maker | Preference Δ | W–L–T | Criteria | Cost / buildCost |
|---|---|---|---|---|---|
| deepseek-v4-pro PreliminaryDeepSeek · —/12 criteria | DeepSeek | +121 vote1 voter · 1 build | 1–0–0 | —/12 | $0.98 |
| claude-sonnet-5 PreliminaryAnthropic · 12/12 criteria | Anthropic | +121 vote1 voter · 1 build | 1–0–0 | 12/12 | $2.22 |
| gpt-6-astra PreliminaryOpenAI · 12/12 criteria | OpenAI | +121 vote1 voter · 1 build | 1–0–0 | 12/12 | $2.69 |
| deepseek-v4-flash PreliminaryDeepSeek · 10/12 criteria | DeepSeek | even1 vote1 voter · 1 build | 0–0–1 | 10/12 | $0.15 |
| grok-build-0.1 PreliminaryxAI · 11/12 criteria | xAI | even1 vote1 voter · 1 build | 0–0–1 | 11/12 | $1.29 |
| gpt-5.6-luna PreliminaryOpenAI · 11/12 criteria | OpenAI | -121 vote1 voter · 1 build | 0–1–0 | 11/12 | $0.044 |
| grok-4.3 PreliminaryxAI · 11/12 criteria | xAI | -121 vote1 voter · 1 build | 0–1–0 | 11/12 | $0.10 |
| grok-4.6 PreliminaryxAI · 12/12 criteria | xAI | -121 vote1 voter · 1 build | 0–1–0 | 12/12 | $3.34 |
No votes yet: glm-5.2, kimi-k3, gpt-5.6-terra, kimi-k2.7-code, grok-4.5, claude-fable-5, claude-opus-5.
4 votes so far, so treat the order as an early signal rather than a verdict. Preference Δ is Elo relative to its 1500 starting point, not a percentage or a confidence interval. Scores depend on the opponents played and vote order. Cost is the median billed build cost for this benchmark. Only votes from server-issued matchups are included. Earlier votes remain archived.
Compare model results and build costs
Compare criteria met, successful builds, model calls, run counts, and actual generation costs. Values are medians across measured runs, with ranges when runs differ. Models with fewer than three runs are provisional.
Build cost and technical resultsTechnical criteria · current benchmark
Higher means more requirements met. Further left means lower cost. Outlined points offer an observed cost–score tradeoff. Small samples are provisional.
kimi-k3 · $0.50 per build · 12/12 criteria · 1 run · 7 min median generation · Provisional
These are technical checks, not player enjoyment or a promise of reliability. Costs summarize completed runs. Failed attempts and run-to-run ranges appear in the table.
No model has reached 3 runs yet, so nothing is ranked. Everything measured so far is below.
| Model | Maker | Criteria | Boots | Cost | Calls | Time | Runs | ||
|---|---|---|---|---|---|---|---|---|---|
| · | kimi-k3 provisionalMoonshot · 100% boots · 29 calls · n=1 | Moonshot | 12/12 | 100% | $0.50 | 29 | 446s | 1 | locked |
| · | claude-sonnet-5 provisionalAnthropic · 100% boots · 51.5 calls · n=2 | Anthropic | 12/12 | 100% | $2.22$1.83–$2.60 | 51.5 | 892s | 2 | locked |
| · | gpt-6-astra provisionalOpenAI · 100% boots · 32 calls · n=2 | OpenAI | 12/12 | 100% | $2.69$2.62–$2.77 | 32 | 386s | 2 | locked |
| · | claude-fable-5 provisionalAnthropic · 100% boots · 27 calls · n=2 | Anthropic | 12/12 | 100% | $3.21$2.87–$3.55 | 27 | 516s | 2 | locked |
| · | grok-4.6 provisionalxAI · 100% boots · 73.5 calls · n=2 | xAI | 12/12 | 100% | $3.34$2.90–$3.78 | 73.5 | 5662s | 2 | locked |
| · | claude-opus-5 provisionalAnthropic · 100% boots · 55.5 calls · n=2 | Anthropic | 11.5/1211–12 | 100% | $3.75$3.65–$3.85 | 55.5 | 815s | 2 | locked |
| · | gpt-5.6-luna provisionalOpenAI · 100% boots · 21 calls · n=1 | OpenAI | 11/12 | 100% | $0.044 | 21 | 230s | 1 | locked |
| · | grok-4.3 provisionalxAI · 100% boots · 12 calls · n=2 | xAI | 11/12 | 100% | $0.10$0.080–$0.12 | 12 | 132s | 2 | locked |
| · | glm-5.2 provisionalZ.ai · 100% boots · 53 calls · n=1 | Z.ai | 11/12 | 100% | $0.49 | 53 | 1692s | 1 | locked |
| · | gpt-5.6-terra provisionalOpenAI · 100% boots · 22.5 calls · n=2 | OpenAI | 11/12 | 100% | $0.53$0.44–$0.61 | 22.5 | 252s | 2 | locked |
| · | kimi-k2.7-code provisionalMoonshot · 100% boots · 51 calls · n=1 | Moonshot | 11/12 | 100% | $0.54 | 51 | 2130s | 1 | locked |
| · | grok-build-0.1 provisionalxAI · 100% boots · 119 calls · n=1 | xAI | 11/12 | 100% | $1.29 | 119 | 1309s | 1 | locked |
| · | deepseek-v4-flash provisionalDeepSeek · 100% boots · 89 calls · n=1 | DeepSeek | 10/12 | 100% | $0.15 | 89 | 6029s | 1 | locked |
| · | grok-4.5 provisionalxAI · 50% boots · 57 calls · n=2 | xAI | 6/120–12 | 50% | $2.12$1.50–$2.74 | 57 | 5280s | 2 | locked |
| · | deepseek-v4-pro provisionalDeepSeek · 100% boots · 86 calls · n=1 | DeepSeek | —/12 | 100% | $0.98 | 86 | 2448s | 1 | locked |
Sign in and judge a matchup to open builds for play. Medians, with the run-to-run range beneath where runs differed. Under 3 runs is provisional. Brief version d13e221fafc3. Raw data: /arena/runs.
Published results for every category
Snapshot from . These tables are available without JavaScript. The live tables above refresh when you open the page.
Arcade results · 15 models · 4 votes
Median results across measured runs. Fewer than 3 runs is provisional. Player preferences are preliminary. Elo starts at 1500; unrated models have no recorded player votes in this snapshot.
| Model | Player Elo (preliminary) | Criteria | Build cost (USD) | Runs |
|---|---|---|---|---|
| kimi-k3 | Not yet rated | 12/12 | $0.50 | 1 (provisional) |
| claude-sonnet-5 | 1512 (1 votes) | 12/12 | $2.22 | 2 (provisional) |
| gpt-6-astra | 1512 (1 votes) | 12/12 | $2.69 | 2 (provisional) |
| claude-fable-5 | Not yet rated | 12/12 | $3.21 | 2 (provisional) |
| grok-4.6 | 1488 (1 votes) | 12/12 | $3.34 | 2 (provisional) |
| claude-opus-5 | Not yet rated | 11.5/12 | $3.75 | 2 (provisional) |
| gpt-5.6-luna | 1488 (1 votes) | 11/12 | $0.044 | 1 (provisional) |
| grok-4.3 | 1488 (1 votes) | 11/12 | $0.10 | 2 (provisional) |
| glm-5.2 | Not yet rated | 11/12 | $0.49 | 1 (provisional) |
| gpt-5.6-terra | Not yet rated | 11/12 | $0.53 | 2 (provisional) |
| kimi-k2.7-code | Not yet rated | 11/12 | $0.54 | 1 (provisional) |
| grok-build-0.1 | 1500 (1 votes) | 11/12 | $1.29 | 1 (provisional) |
| deepseek-v4-flash | 1500 (1 votes) | 10/12 | $0.15 | 1 (provisional) |
| grok-4.5 | Not yet rated | 6/12 | $2.12 | 2 (provisional) |
| deepseek-v4-pro | 1512 (1 votes) | —/12 | $0.98 | 1 (provisional) |
Shooter results · 14 models · 25 votes
Median results across measured runs. Fewer than 3 runs is provisional. Player preferences are preliminary. Elo starts at 1500; unrated models have no recorded player votes in this snapshot.
| Model | Player Elo (preliminary) | Criteria | Build cost (USD) | Runs |
|---|---|---|---|---|
| claude-opus-5 | 1544 (4 votes) | 11.5/12 | $5.08 | 2 (provisional) |
| grok-4.5 | 1478 (5 votes) | 10.5/12 | $2.14 | 2 (provisional) |
| deepseek-v4-pro | 1499 (2 votes) | 10/12 | $0.89 | 1 (provisional) |
| deepseek-v4-flash | 1488 (4 votes) | 9.5/12 | $0.092 | 2 (provisional) |
| gpt-5.6-terra | 1489 (3 votes) | 9.5/12 | $0.46 | 2 (provisional) |
| claude-sonnet-5 | 1533 (4 votes) | 9.5/12 | $3.73 | 2 (provisional) |
| grok-build-0.1 | 1501 (2 votes) | 9/12 | $1.14 | 1 (provisional) |
| glm-5.2 | 1468 (4 votes) | 8.5/12 | $0.42 | 2 (provisional) |
| kimi-k3 | 1466 (3 votes) | 8/12 | $0.40 | 1 (provisional) |
| kimi-k2.7-code | 1510 (4 votes) | 8/12 | $0.69 | 1 (provisional) |
| claude-fable-5 | 1536 (3 votes) | 8/12 | $6.67 | 2 (provisional) |
| gpt-6-astra | 1523 (2 votes) | 7/12 | $3.31 | 2 (provisional) |
| grok-4.6 | 1487 (7 votes) | 7/12 | $4.11 | 2 (provisional) |
| grok-4.3 | 1477 (3 votes) | 5/12 | $0.30 | 1 (provisional) |
Puzzle results · 14 models · 9 votes
Median results across measured runs. Fewer than 3 runs is provisional. Player preferences are preliminary. Elo starts at 1500; unrated models have no recorded player votes in this snapshot.
| Model | Player Elo (preliminary) | Criteria | Build cost (USD) | Runs |
|---|---|---|---|---|
| grok-4.6 | 1524 (2 votes) | 11.5/12 | $3.11 | 2 (provisional) |
| claude-opus-5 | 1500 (1 votes) | 11.5/12 | $4.09 | 2 (provisional) |
| gpt-5.6-terra | 1488 (2 votes) | 11/12 | $0.41 | 2 (provisional) |
| kimi-k3 | Not yet rated | 11/12 | $0.53 | 1 (provisional) |
| claude-sonnet-5 | 1500 (1 votes) | 11/12 | $3.89 | 2 (provisional) |
| glm-5.2 | 1489 (3 votes) | 10.5/12 | $0.58 | 2 (provisional) |
| gpt-6-astra | 1500 (1 votes) | 10.5/12 | $1.92 | 2 (provisional) |
| claude-fable-5 | 1500 (1 votes) | 10.5/12 | $6.08 | 2 (provisional) |
| grok-4.3 | 1476 (2 votes) | 10/12 | $0.29 | 2 (provisional) |
| kimi-k2.7-code | 1512 (1 votes) | 10/12 | $0.35 | 1 (provisional) |
| deepseek-v4-flash | 1499 (2 votes) | 9.5/12 | $0.10 | 2 (provisional) |
| grok-4.5 | 1512 (1 votes) | 9/12 | $1.42 | 1 (provisional) |
| grok-build-0.1 | 1500 (1 votes) | 8/12 | $1.47 | 2 (provisional) |
| deepseek-v4-pro | Not yet rated | —/12 | Not available | 0 (provisional) |
RPG results · 14 models · 7 votes
Median results across measured runs. Fewer than 3 runs is provisional. Player preferences are preliminary. Elo starts at 1500; unrated models have no recorded player votes in this snapshot.
| Model | Player Elo (preliminary) | Criteria | Build cost (USD) | Runs |
|---|---|---|---|---|
| grok-4.5 | 1512 (1 votes) | 12/12 | $2.24 | 1 (provisional) |
| claude-sonnet-5 | 1488 (2 votes) | 12/12 | $4.57 | 1 (provisional) |
| claude-opus-5 | 1511 (2 votes) | 12/12 | $5.11 | 1 (provisional) |
| deepseek-v4-flash | 1512 (1 votes) | 11.5/12 | $0.11 | 2 (provisional) |
| gpt-5.6-terra | Not yet rated | 11.5/12 | $0.38 | 2 (provisional) |
| gpt-6-astra | 1524 (2 votes) | 11.5/12 | $2.29 | 2 (provisional) |
| claude-fable-5 | 1488 (1 votes) | 11.5/12 | $5.67 | 2 (provisional) |
| deepseek-v4-pro | Not yet rated | 11/12 | $1.32 | 1 (provisional) |
| kimi-k2.7-code | 1476 (2 votes) | 10/12 | $0.77 | 1 (provisional) |
| grok-4.6 | Not yet rated | 10/12 | $4.64 | 1 (provisional) |
| grok-4.3 | 1488 (1 votes) | 9.5/12 | $0.40 | 2 (provisional) |
| grok-build-0.1 | 1488 (1 votes) | 6.5/12 | $1.18 | 2 (provisional) |
| glm-5.2 | 1512 (1 votes) | 0/12 | $0.60 | 2 (provisional) |
| kimi-k3 | Not yet rated | —/12 | Not available | 0 (provisional) |
Community challenges
What should the models build next?
Propose a game idea, refine the brief together, and vote for your favorites.
Loading community challenges…
How the comparison works
Every model gets the same three-message brief within its category. We balance model exposure and sample an eligible build, so a model’s best attempt does not stand in for every run. Matchups, ratings, and costs use the same brief version, worker version, and prompt variant.
A signed-in account can judge each build once. Accepted votes cannot be changed after reveal. “Both need work” records a tie and a separate quality flag. Skips and broken-game reports do not affect ratings. Earlier votes collected without server-issued matchups remain in the archive and are excluded from the new preference table.
Model identities are hidden in the comparison interface. Public run records remain available for inspection, so someone familiar with a published game may recognize it. These results measure preference within Cinevva’s game-creation workflow.
Game categories
Choose Arcade for waves, lives, and scoring; Shooter for 3D camera control, aiming, and hitscan; Puzzle for grids, matches, and cascades; or RPG for inventory, combat, and saved progress. The brief and ratings are specific to each category.
Arcade benchmark brief and scoring
The example below is the Arcade brief. Shooter, Puzzle, and RPG use their own briefs; switching the category changes the matchups and results above.
The three messages every model gets, word for word
1 — Build it
Build me a vertical space shooter arcade game. The player controls a ship at the bottom of the screen that moves left and right with the arrow keys and also A and D, and fires bullets with the spacebar with a sensible fire-rate limit. Enemies spawn in waves at the top and move downward, and the waves get harder over time by spawning faster or moving faster. A bullet hitting an enemy destroys both and adds to the score. The score must be visible on screen at all times, and so must a life counter starting at 3 lives, where a life is lost when an enemy reaches the bottom or touches the ship. At zero lives show a clear game over message, and pressing R restarts the game without reloading the page. It has to be playable on a phone at 375px wide with on-screen touch controls, and expose window.getGameState() returning at least the score and the lives remaining.
2 — Extend it
Add a power-up system. Every so often a destroyed enemy should drop a power-up that falls toward the player, and catching it gives a temporary effect that ends after a few seconds. Include at least two different kinds, one that changes how the ship shoots and one that protects the ship. Make it obvious on screen which power-up is active and how long is left, and include the active power-up in what window.getGameState() returns.
3 — Break the news
I tried it on my phone and the buttons don't do anything. I can see them but tapping does nothing, the ship just sits there. Also it feels like the whole thing is too zoomed in on a small screen.
Message three is the one that separates models. It's vague, it's from a phone, and the fault it describes might not even exist in that model's build. Saying "that already works, here's why" is a pass. Announcing a fix you didn't make is not.
The twelve criteria, and how they're scored
- Ship moves with arrow keys and A/D
- Spacebar fires, rate-limited
- Enemies spawn in waves and descend
- Bullet hits enemy: both die, score goes up
- Score always visible
- 3 lives, visible, lost correctly
- Game-over state with a message
- R restarts without a page reload
- Difficulty climbs over time
- Playable at 375px with working touch controls
window.getGameState()returns score and lives- No uncaught console errors during play
Two judges from different AI companies score every build independently. A criterion counts only when both agree. Splits are marked contested, not quietly resolved.
Boots is mechanical, not judged: the JavaScript is parsed, and a file that won't parse fails however good it looks. Cost is what the run actually billed, read from our provider records, covering the model's whole tool loop.
How we keep it fair
A run counts only if the model we asked for is the model that billed — our system can fall back to another provider mid-run, and we're not publishing one company's work under another's name.
Medians with ranges, never averages. n on every card. Under three runs is marked provisional.
Results never pool across a change to our prompt.
These are results through our harness — our brief, our system prompt, our tools. A model that scores modestly here may shine in a different setup.
Every build here was made in the game creator with the picker set to that model. Run the same brief yourself and see what you get.
About the AI game benchmark
Which AI model is best for making games?
There is no single winner across all games. Use the category ratings to compare player preferences, then check the technical results, costs, and number of runs. These results measure models through Cinevva's game-creation tools and prompts; they are not a universal ranking of coding ability.
How does blind voting work?
Play both builds before choosing a winner or a tie. Model names and individual build costs appear after a vote is recorded. Voting requires a registered account, and the Arena avoids showing you builds you have already judged.
What does the build cost include?
The cost beneath a revealed game is the amount billed for that specific build's model tool loop. The comparison table shows the median across a model's measured runs. This is the generation cost, not a price to play or fork the game.
Can I fork an Arena game and edit it?
After voting, choose “Edit this game” on either revealed build to create your own editable copy in Game Creator. The editor opens automatically when the copy is ready. Forking is available for games whose creators allow it.
Can I play on a phone?
Yes. The Arena provides a mobile player, and the Arcade brief explicitly asks models for working touch controls. Touch support is also a scored criterion, so individual AI-generated builds may still have control or layout problems.
Build your own browser game
Try the same brief in Cinevva Game Creator, browse the browser game arcade, or use the game prompt guide to write a different challenge. Learn more about the workflow in vibe coding games.