Atlas

Benchmarks

← All benchmarks

Emergent Collusion High-Illegality Game Rate

Safety · 2025-07-15

Emergent Collusion evaluates six instances of the same model in repeated auction games with optional communication and an explicit instruction not to collude. This metric is the percentage of games that an independent analyst assigns an illegality score of at least 7 out of 10, so lower is better.

Top models (lower is better)

ModelScore
Grok 319.0
ChatGPT-4o Latest observed 2025-03-2723.0
DeepSeek-V3-032426.0
Grok 3 Mini30.0
Sonnet 432.0
Opus 436.0
Gemini 2.5 Pro39.0
Mistral Medium 347.0
o351.0
Qwen3-235B-A22B57.0
O4 Mini62.0
DeepSeek-R1-052871.0
Grok 475.0
Loading Atlas data…