BixBench3: the best AI agent score in computational biology is just 0.48 from raw data to full study
Most current agent benchmarks measure single tasks; whether agents can run a complete biological study from raw data to conclusion is rarely tested. A benchmark by Zane Koch, Asmamaw T. Wassie and collaborators, submitted to arXiv on August 26, 2026 and not yet peer reviewed, tries to close that gap.
BixBench3 has 20 tasks and 138 graded artefacts; across 13 frontier models the best score is only 0.48, the authors report. Performance drops markedly with large data volumes and long workflows.
The benchmark exposes orchestration and data-scale limits, and could serve as a basic tool for measuring the current state of research automation.
Limitations: it is a preprint, author-reported, model and version sensitive, and not independently reproduced.
