AI Models Struggle on New Biology Research Benchmark
PhD biologists built a 50-task benchmark in which five frontier AI models succeeded on only 4% to 18% of tasks.
Updated on Nov. 3, 2026 in Life Sciences
Live Poll
Would you trust an AI assistant to help you interpret important test results?
PhD biologists developed BioResearchBench, a benchmark of 50 biology research tasks. Five frontier AI models were evaluated on it, with whole-task success rates ranging from 4% to 18%.
Why it matters
The results suggest that current AI systems remain far from reliably carrying out end-to-end biological research workflows. A benchmark grounded in real sequencing, proteomic, microscopy and clinical data offers a common yardstick for measuring progress worldwide.
The benchmark covers 50 research tasks across 10 biology subdomains, drawing on sequencing data, proteomic measurements, microscopy images, assay readouts and clinical datasets. Failure rates were highest for computation and data handling (66.9%), followed by uncertainty and identification (55.3%) and reporting and completeness (54.7%).
The players
BioResearchBench
A benchmark of 50 biology research tasks developed by PhD biologists to test whether AI models can carry out research-level biological work.
Claude Opus 5.5
An AI model that achieved the highest whole-task success rate of the five frontier models evaluated, at 18%.
Gemini 4 Argon
An AI model that achieved 13% whole-task success on BioResearchBench.
GPT-6 Astra
An AI model that achieved 4% whole-task success on BioResearchBench.
The details
The 50 research questions were written by PhD biologists and refined through expert review and iterative model testing. Model responses were scored against rubric requirements, with failure rates calculated by category across the proportion of eligible tasks.
Timeline
3 November 2026: BioResearchBench preprint published on bioRxiv.
The Big Picture
Most AI benchmarks to date have measured general knowledge or coding ability, so BioResearchBench shifts evaluation toward multi-step scientific work. Its low success rates indicate that the gap between answering biology questions and conducting biology research remains wide.
For researchers and labs considering AI tools, the results indicate that current models will need substantial human oversight on multi-step biology work. The benchmark's public release would let institutions test future models against the same tasks before adopting them.
The takeaway
Before relying on AI for biological research, check how it performs on tasks that match your own workflow, since success rates in this benchmark ranged only from 4% to 18%. Human review remains essential at each stage.
Further reading
For related coverage, see Life Sciences on Wisevoter.
Source note: This article includes information reported by Biorxiv.
Live Poll
Would you trust an AI assistant to help you interpret important test results?
