OpenAI Ships BioBasicsBench with Ligo: 100 Routine Wet-Lab Tasks; Vendor GPT-6 Astra ~66.7% at xhigh

Core Background & Industry Pain Points

Everyday wet-lab work is full of failed runs, ambiguous images, and conflicting observations. Many biology benchmarks use paper-ready charts and miss reading raw evidence, checking controls, and saying when data cannot support a conclusion. On 2026-11-06, OpenAI and enzyme-design biotech Ligo introduced BioBasicsBench: 100 routine wet-lab tasks from real experiments for interpreting raw data, spotting errors, and withholding unsupported conclusions.

Architecture Highlights & Internals

Tasks span protein methods, imaging/cytometry, quantification/assays, nucleic-acid methods, and molecular cloning. Inputs keep gels/blots, plate photos, microscopy stacks, sequencing FASTQs, plate-reader exports, and bench videos, plus ladder/plasmid reference sheets. Each item has an expert rubric and passed Ligo peer review, an external PhD wet-lab review, and a final internal pass. Agents return structured answers scored criterion-by-criterion.

Authoritative Benchmarks & Measured Scores

Figures below are OpenAI-reported at xhigh (100 tasks × 5 attempts), not reproduced here. GPT-6 Astra mean reward ~66.7% (95% CI 60.3–73.1%); GPT-6.1 Sol ~62.0%; GPT-5.6 Sol ~43.8%. OpenAI notes Astra also leads at lower effort, while still failing on mislabeled controls, assumption-over-evidence, and colony counting.

Developer Hands-on Guide

Read the official post before product claims. For life-science agents, add regressions for misread gels, ignored controls, and over-interpretation; verify antibiotic labels and channel crosstalk. The bench tracks routine analysis and biosafety implications—it does not claim autonomous lab judgment; keep humans in the loop.