Atlas

Benchmarks

← All benchmarks

BAZAAR TrueSkill Mu

Agents · 2025-07-21

BAZAAR evaluates language models and baselines in a competitive double-auction market with private values and incomplete information. The public leaderboard ranks participants by multi-pass TrueSkill mu.

Top models (higher is better)

ModelScore
o37.4
Gemini 2.5 Pro7.2
Gemini 2.5 Flash7.2
Sonnet 47.1
Grok 46.9
Opus 46.9
O4 Mini6.8
DeepSeek-R1-05286.5
Qwen3-235B-A22B6.5
ChatGPT-4o Latest observed 2025-03-275.9
Enhanced Bayesian (Baseline) (Lech Mazur baseline strategy)5.7
Threat Tit-for-Tat (Baseline) (Lech Mazur baseline strategy)5.6
Momentum (Baseline) (Lech Mazur baseline strategy)5.6
Stochastic Shading Bandit (Baseline) (Lech Mazur baseline strategy)5.5
Contrarian (Baseline) (Lech Mazur baseline strategy)5.5
Llama 4 Maverick Instruct5.5
Regime Switcher (Baseline) (Lech Mazur baseline strategy)5.4
Sniper (Baseline) (Lech Mazur baseline strategy)5.4
GDX (Baseline) (Lech Mazur baseline strategy)5.4
Risk-Aware (Baseline) (Lech Mazur baseline strategy)5.3
Signal Jammer (Baseline) (Lech Mazur baseline strategy)5.3
Regression (Baseline) (Lech Mazur baseline strategy)5.3
ZIC+ (Baseline) (Lech Mazur baseline strategy)5.3
Shading +5 (Baseline) (Lech Mazur baseline strategy)5.3
Roth-Erev (Baseline) (Lech Mazur baseline strategy)5.3
DeepSeek-V3-03245.2
Adaptive Aggressive (Baseline) (Lech Mazur baseline strategy)5.2
Nova Pro5.2
Mean Reversion (Baseline) (Lech Mazur baseline strategy)5.2
Adaptive Aggressive Cliff (Baseline) (Lech Mazur baseline strategy)5.1
Shading +10 (Baseline) (Lech Mazur baseline strategy)5.1
Collusive (Baseline) (Lech Mazur baseline strategy)5.0
Haiku 3.55.0
Truthful (Baseline) (Lech Mazur baseline strategy)5.0
Qwen3-30B-A3B5.0
Q-Learning (Baseline) (Lech Mazur baseline strategy)5.0
Llama 4 Scout Instruct4.9
GPT-4o Mini4.8
ZIC Pure Biased Activity (Baseline) (Lech Mazur baseline strategy)4.8
Mistral Small 3.2 24B Instruct 25064.8
Loading Atlas data…