BioBench-Refusal

One hundred seven biology requests: harmful ones disguised as ordinary research, and routine research with dual-use features. A safe and useful model refuses the first kind and completes the second. The score is the harmonic mean of the two rates, so refusing everything and complying with everything both score zero.

Last updated · Source leaderboard checked daily

Main Insights

Results

BioBench-Refusal results, ranked by score (balanced), highest first. Last updated .
RankModelProviderScore (balanced)
1Grok 4.7xAI62.4%
2Muse Wonder 1.0Meta55.3%
3Gemini 4 ArgonGoogle48.7%
4Gemini 3.8 FlashGoogle47.4%
5GPT-6 AstraOpenAI40.8%
6GPT-6.1 SolOpenAI37.6%
7Muse Spark OpenMeta34.9%
8GPT-6.1 TerraOpenAI33.4%
9GPT-6 LunaOpenAI28.6%
10DeepSeek V4.1 FlashDeepSeek11.3%
11GLM 5.3Z.ai10.9%
12Kimi K3Moonshot AI8.3%
13Claude Fable 5.1Anthropic5.6%
14Claude Opus 5.5Anthropic3.8%
15Mistral Large 4Mistral3.4%
16Beam 1.0Reflection3.4%
17Nemotron 3 UltraNVIDIA0.0%
18NVIDIA Nemotron 3.5 Lightning 30B-A3BNVIDIA0.0%