Output Loop Bench · Artifacts

Simulation: AI models compared

Physics and motion that must behave correctly. 29 models tested on the same task.

Balls in a rotating hexagon

The best-known public physics test. Weak models let balls pass through the walls; strong ones get collisions and the moving walls right.

ModelMakerTimeTokens
gemini-3.1-pro ★ Output Loop's pickGoogle207.6s22,718
muse-spark-1.3-mediumMeta31.4s27,819
composer-2.5Cursor33.9s25,421
gemini-4-argon-mediumGoogle42.2s20,475
nemotron-3-nano-4bNVIDIA45.1s4,592
MiniMax-M2.5MiniMax45.8s18,920
kimi-k2.7-codeMoonshot (Kimi)53.3s35,182
kimi-k3-maxMoonshot (Kimi)59.4s27,525
claude-sonnet-5-5Anthropic60.6s97,609
bonsai-27bPrism ML69.2s4,016
gpt-6-astraOpenAI81.4s17,642
claude-opus-5-5Anthropic103.9s98,714
gemma-4-26b-a4b-qatGoogle115.5s13,434
claude-haiku-5-5Anthropic120.6s114,838
gpt-6.1-solOpenAI126.3s19,161
qwen3.8-flashAlibaba (Qwen)134.8s23,646
glm-5.3Z.ai138.0s48,420
claude-fable-5-1Anthropic139.9s102,746
gemini-3.8-flashGoogle154.3s34,611
nemotron-3.5-lightning-30b-a3bNVIDIA169.1s33,530
deepseek-v4-proDeepSeek176.6s39,503
deepseek-v4.1-flashDeepSeek224.6s38,524
muse-glimmerMeta319.9s10,045
grok-4.7-build-fastxAI393.4s63,320
gpt-6-lunaOpenAI396.0s35,480
kimi-k3Moonshot (Kimi)453.2s23,419
grok-4.7xAI472.5s53,864
qwen3.8-maxAlibaba (Qwen)477.4s34,357
muse-glimmer-30bMeta1083.9s23,614

Open Simulation in the interactive Bench