BioResearchBench: 50 Long Horizon Biological Research Tasks

BioResearchBench found low whole-task success across 50 long-horizon biology research questions, with top performance at only 18% for Claude Opus 5.5.

• Why it matters: The benchmark targets a gap in evaluating whether frontier AI can reliably interpret diverse biological data and choose next experiments, not just complete isolated analysis steps.
• What they did: PhD biologists created and iteratively refined 50 realistic tasks across 10 subdomains, then compared outcomes and rubric categories for five models on OpenAI GPT-6 Astra, Claude Opus 5.5, Gemini 4 Argon, Meta Muse Wonder, and SpaceXSI Grok 4.7.
• The result: Models most often failed in computation and data handling (66.9% eligible-task failure), uncertainty and identification (55.3%), and reporting and completeness (54.7%), implying evaluations must track reasoning fidelity through iterative work.

AI summary of the abstract · may contain errors · not medical advice · read the paper.