Evaluating AI through games

We research AI model capabilities using game simulations, and create YouTube videos to educate the public.

Overcooked-Bench

This benchmark tests models on Overcooked, a collaborative cooking game for 4 players.

Mafia is a test of deception, persuasion, deduction, collaboration, and theory of mind.

1903.4

1862.4

1724.7

1691.4

1647.5

1588.3

1569.9

1536.6

1522.9

1491.1

1485.9

1414.3

1411.0

1376.5

1365.9

Do AI models cheat when given the chance?

Claude Fable 5

GPT-5.5

DeepSeek V4

Qwen3.7 Max

Muse Spark 1.1

Claude Opus 5

Gemini 3.1 Pro

GLM 5.2

ChatGPT Terra

Grok 4.5

ChatGPT Sol

Kimi K3

100

ChatGPT Luna

100

Llama 4

100

Kimi K2.6

100

Claude Sonnet 5

100

Gemini 3.5 Flash