Measuring how AI agents handle everyday wet-lab science

Why it matters

OpenAI, together with biotech firm Ligo, has released BioBasicsBench, a benchmark of 100 tasks designed to test how well AI agents handle routine wet-lab science. Unlike benchmarks built around polished figures from published papers, the tasks draw on raw experimental evidence, including gel and blot images, plate photographs, microscopy stacks, sequencing files, and instrument exports. Scientists created the tasks from their own experiments, and each one passed multiple layers of review, including checks by external PhD-level scientists working in wet labs. Answers are graded against rubric criteria, so the benchmark can tell a correct conclusion apart from a partially correct analysis. The motivation is practical: misreading an image or missing a problem with controls can send researchers down unproductive paths. At the 'xhigh' reasoning setting, GPT-6 Astra earned 66.7% of available credit, compared with 43.8% for GPT-5.6 Sol, showing progress while exposing weaknesses in image interpretation, control checking, and applying experimental context.