Atlas

Benchmarks

← All benchmarks

BioMysteryBench - Human Difficult

Science · 2026-07-24

BioMysteryBench Human Difficult contains problems that remain unsolved by independent human experts but have objective ground-truth solutions. This is the revised subset reported from the Claude Opus 5 system card onward, after Anthropic removed 6 problems on external feedback; earlier original-set results use biomysterybench-human-difficult-original-set.

Top models (higher is better)

ModelScore
Gemini 3.8 Flash56.5
Claude Opus 549.4
Claude Fable 546.5
Gemini 3.7 Flash43.5
Opus 4.842.4
Sonnet 534.1
Loading Atlas data…