Most science benchmarks for AI use figures cleaned up for publication. Real lab work looks different: smudged gels, mislabelled plates, out-of-focus microscope stacks and controls that quietly fail. OpenAI and Ligo, a biotech company that designs enzymes, have built BioBasicsBench to test agents on that messier evidence, and the results show both real progress and the kind of errors that would cost a lab a week.
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. Scientists wrote them from their own experiments, using gel and blot images, plate photos, raw microscopy, sequencing files, plate-reader exports and even videos of bench work. Each task went through three rounds of review, including one by an outside PhD-level scientist, and is graded against an expert rubric that gives partial credit and rewards saying when the evidence is not enough.
OpenAI tested seven of its models across 37 reasoning settings, five attempts per task, for 18,500 results. At the highest setting common to all, GPT-6 Astra earned 66.7% of available credit, ahead of GPT-6.1 Sol at 62.0% and well ahead of GPT-5.6 Sol at 43.8%, while using about half as many output tokens per task. The examples are telling. Astra noticed that a negative-control plate was labelled for kanamycin when the experiment used ampicillin, and rejected it; it recognised a series of microscope images as a depth scan rather than cells vanishing; and it declined to call cells dead from a dye signal that a second dye could have produced.
For readers of our GPT-6 Astra review, this supports the claim that its gains come largely from stronger handling of images and other raw data. OpenAI's own conclusion points the same way, crediting better native multimodal ability.
The failures are as instructive. Astra let a familiar assumption, that a larger PCR product means successful cloning, override what the controls showed, and picked the wrong colonies. It discounted lighter vacuoles that contradicted its preferred description, and overcounted colonies on one plate by about 12%. A third of the credit was still missed on routine work, so any lab using such an agent should keep a scientist checking each reading. The benchmark is OpenAI's own, built with one partner and tested only on OpenAI models, so how rival systems compare is not yet known.