BehaviorBench - Beauty Contest Win Rate: leaderboard
BehaviorBench (University of Michigan, 2026) evaluates foundation models on behavioral science tasks at the individual and population levels, using economic-game records, Big Five personality surveys and research literature. This board is the win rate (%) in the beauty contest game against recorded human players, where the model must guess closest to two thirds of the group average; higher is better.
Metric: Win rate (%) in the beauty contest game played against recorded human players: share of sessions in which the guess of the model is closer to two thirds of the group average than every human guess; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 17 | #161 |
| 2 | GPT-4.1 | 17 | #495 |
| 3 | Claude Opus 4.6 | 15 | #102 |
| 4 | DeepSeek V3.2 | 14 | #406 |
| 5 | Gemini 3.1 Flash Lite (Preview) | 14 | #418 |
| 6 | Claude Haiku 4.5 | 13 | #568 |
| 7 | Gemini 3.1 Pro (Preview) | 12 | #94 |
| 8 | GPT-5.4 Mini (High) | 10 | #389 (GPT-5.4 Mini) |
| 9 | Qwen 3 4B | 9 | #2281 |
| 10 | GPT-5.4 (High) | 8 | #123 (GPT-5.4) |
| 11 | Llama 3.3 70B Instruct | 6 | #1357 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=behaviorbench-beauty-contest-win-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.