OpenAI published BioBasicsBench on November 6, 2026, a benchmark developed with biotech company Ligo to test how AI agents interpret raw evidence from everyday wet-lab experiments. Its 100 tasks ask for conclusions, measurements, error detection and sometimes a recognition that the available evidence is insufficient. This is a measurement of model answers to prepared tasks, not a demonstration that an agent can independently run a laboratory or safely make medical decisions. VIBPS has not reproduced the evaluation.
What does BioBasicsBench test?
The tasks draw on material scientists encounter before results are turned into polished figures: gel and blot images, colony plates, microscope views, sequencing files, instrument exports, laboratory notes and videos of procedures. OpenAI groups them across protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. One task may involve several kinds of evidence. Scientists wrote grading criteria for each question; OpenAI says a different Ligo scientist, an outside wet-lab scientist and an internal reviewer checked each task. Those review steps improve the benchmark's design, but the published score remains from a benchmark developed with OpenAI, not an independent clinical or field validation.
An agent returns a structured answer and, where requested, supporting reasoning. A rubric gives partial credit for individual criteria rather than treating every answer as simply right or wrong. That makes the reported percentage a mean share of available rubric credit. It is not the percentage of experiments an agent could conduct successfully, nor the chance that a scientific conclusion is safe to use without review.
How did the tested models score?
At the xhigh reasoning setting, OpenAI reports a 66.7% mean reward for GPT-6 Astra and 43.8% for GPT-5.6 Sol, with five attempts per task. The same table gives GPT-6.1 Sol 62.0%. These are comparisons within the benchmark's 100 selected tasks, under the stated model and reasoning settings. The publication does not show that a lab using an agent improves its research output by the difference between those scores. It also reports uncertainty intervals, so small apparent gaps should not be read as decisive without the full evaluation design.
Where do agents still get the science wrong?
OpenAI includes failure cases for its strongest model. In one cloning task, Astra grouped bands in a gel but associated the groups with the wrong experimental identities, leading it to nominate the wrong colonies for follow-up. In another case it noticed that microscope contrast varied with focus but gave a final answer that ignored contradictory lighter examples. On a colony-counting task, the model's marks missed visible colonies, counted empty areas and produced an overcount of about 12% against the benchmark reference. These examples show why a plausible interpretation and a high aggregate score cannot replace checking controls and raw observations.
Why should software makers pay attention?
The useful software question is how an AI app reports uncertainty when the input is messy and a wrong answer matters. BioBasicsBench rewards noticing insufficient evidence, not only producing fluent explanations. An app that helps researchers inspect data would need to preserve the original evidence, show which control supports a claim and allow a domain expert to challenge the conclusion. That is VIBPS's inference from the benchmark, not an assertion that a specific research app already implements it. Separate VIBPS coverage of Anthropic's Claude research-use study asks what researchers are doing with AI; this release asks how well agents interpret a defined set of raw lab tasks.