Atlas

Benchmarks

← All benchmarks

BioMysteryBench — Hard (Vals)

Science · 2026-09-22

The 17 human-difficult bioinformatics tasks, with three Terminus 2 trials per task.

Top models (higher is better)

ModelScore
Claude Opus 543.1
Kimi K341.2
Gemini 4 Argon39.2
GPT-6.1 Sol39.2
Hy4 Preview39.2
GPT-6 Astra37.3
GPT-6 Sol35.3
Muse Wonder Full33.3
Claude Haiku 5.529.4
Ember-129.4
GPT-5.6 Sol29.4
Grok 4.629.4
Grok 4.729.4
Kimi K3.1 Pro29.4
MiMo-V2.6-Flash29.4
Mistral Large 4 Preview29.4
Muse Wonder29.4
Step 5 Preview29.4
Gemini 3.1 Pro Preview27.5
GPT-5.6 Luna27.5
Muse Spark 1.225.5
DeepSeek-V4-Flash-073123.5
DeepSeek V4.1 Flash21.6
Gemini 3.6 Flash21.6
Gemini 3.8 Flash21.6
GPT-6 Luna17.6
Beam 1.0 501B A23B15.7
Loading Atlas data…