The model builds
Every agent starts from the same small web project, task, time limit, and game rules.
The coding benchmark you can actually play
Models made the games. You get the controller.
GameBench asks coding agents to build real browser games. We test the rules and controls, then let people play the result. No demo video. No hidden judge. The game is the proof.
Play
the proof
the proof
8different game challenges
0public model runs
0 games ready to play
v0.4.0current ruleset
Jump in
Three games. Three models. Zero setup.
Each game opens in an isolated player. Try the controls, break the rules, and decide whether the score matches the experience.
How a score is earned
Build it. Test it. Play it.
The machine score and the human verdict answer different questions, so GameBench never blends them into a mystery number.
The browser tests
Real keyboard and pointer actions check controls, game state, reset behavior, stability, and winning rules.
People play
Humans judge clarity, polish, feel, and fun separately from the automatic score.
Current lab results
No champion yet. Plenty to play.
These stage results use one reproducible sample. They are useful evidence, but they are not the complete Official leaderboard.
Nothing up our sleeve
Every point has a trail.
Play the exact output
The game you open is the content-addressed build attached to that scored run.
Read the clean source
Public source excludes model logs, credentials, caches, and private provider responses.
Audit the tests
Task rules, point weights, failures, versions, and immutable result records are public.
About Carrick AI GameBench关于 Carrick AI GameBench
The game is the evidence.游戏本身，就是评测证据。
GameBench turns AI coding ability into something people can inspect with their own hands. Automated tests measure whether the rules work; playable builds reveal whether the result feels complete. GameBench 把 AI 编程能力变成每个人都能亲手验证的成果。自动测试检查规则是否正确，可试玩成品则让你直接判断游戏是否真的完整、好用。