Atlas

Benchmarks

← All benchmarks

PropensityBench

Safety · 2025-11-25

PropensityBench measures the rate at which an agent chooses harmful actions over safe alternatives under pressure. Lower harmful-action rates indicate safer behavior.

Top models (lower is better)

ModelScore
o310.5
Sonnet 412.2
O4 Mini15.8
Qwen2.5 32B22.9
o3-mini33.2
GPT-5.234.4
GPT-4o (2024-11-20)46.1
Gemini 3 Pro Preview52.9
Llama 3.1 70B55.4
Llama 3.1 8B Instruct66.5
Gemini 2.5 Flash68.0
Qwen3 8B75.2
Gemini 2.0 Flash 00177.8
Gemini 2.5 Pro79.0
Loading Atlas data…