All benchmarks

BioMysteryBench (73/17 split)

Biology

BioMysteryBench presents unsolved-style biology puzzles that require generating and testing hypotheses against experimental evidence, such as identifying a knocked-out gene from transcriptomic data. The revised task set splits into two subsets: the 73 problems independent human experts solved, and the 17 that remain unsolved by humans but have an objective ground-truth answer.

Model scores

  • Fable 5.2—
  • Mythos 5.2—
  • Fable 5.1—
  • Fable 5—
  • Opus 5.589.3% (human solvable) / 50.0% (human difficult)
  • Opus 591.4% (human solvable) / 51.8% (human difficult)
  • Opus 4.8—
  • Sonnet 5.5—
  • Sonnet 584.9% (human solvable) / 39.4% (human difficult)
  • GPT-6 Astra—
  • GPT-6.1 Sol—
  • GPT-5.6 Sol—
  • Grok 4.7—
  • GPT-5.5—
  • Composer 2.5—
  • Opus 4.7—
  • Gemini 3.1 Pro—
  • Mythos Preview90.3% (human solvable) / 44.1% (human difficult)

Official source: Anthropic — Claude Opus 5.5 (revised BioMysteryBench split)

Related reading