OpenAI's BioBasicsBench tests agents on 100 raw wet-lab tasks, and GPT-6 Astra scores 66.7%
Victoria Shinder1 min readOpenAI has published BioBasicsBench, a benchmark that measures how well AI agents interpret the raw evidence of everyday laboratory biology. The post, dated 6 November, argues that existing benchmarks rely on charts and images selected for scientific papers, while real wet-lab work involves failed experiments, ambiguous images and conflicting observations.
The benchmark has 100 tasks across five areas of routine work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Inputs keep the form scientists actually handle: gel and blot images, plate photographs, microscopy stacks, sequencing reads, plate-reader exports and videos of bench work. Scientists built the tasks from their own experiments and wrote a grading rubric for each; every task was reviewed internally by a scientist who had not written it, externally by a PhD-level wet-lab scientist, and once more internally.
Answers are scored per rubric criterion, so a correct conclusion is separated from a partially correct analysis, and some tasks reward recognising that the evidence is insufficient. OpenAI evaluated seven models across 37 model and reasoning-effort settings, with five attempts per task, for 18,500 results. At the xhigh setting GPT-6 Astra earned 66.7% of the available credit, against 43.8% for GPT-5.6 Sol. The company says the gain holds even at low reasoning effort and suggests stronger native multimodal capability is the likely reason.
The post also says where models still fall short: interpreting images, checking the controls that validate an experiment and applying experimental context.
What it means
A score of two thirds on routine tasks is progress and also a warning. Misreading a gel or missing a broken control is exactly the error that sends a lab down the wrong path, so the benchmark's credit for saying the data cannot support a conclusion is the part worth watching.