OpenAI introduced BioBasicsBench on November 6, a benchmark developed with Ligo to test routine wet-lab analysis. Its research report focuses on what models can infer from raw experimental data, including when a control or an ambiguous image undermines the proposed conclusion.
What BioBasicsBench asks models to do
The benchmark contains 100 tasks across five areas of lab work, including imaging, assays and molecular cloning. Scientists used data from their own experiments and supplied rubrics that award credit for individual parts of an answer. The task may require a conclusion, a measurement or a recognition that the supplied evidence cannot resolve the question.
This design gives a model credit for partial work. A reader should interpret mean reward as the average share of available rubric credit.
How OpenAI compared the models
OpenAI reports that GPT-6 Astra scored 66.7% at the xhigh reasoning-effort setting. GPT-5.6 Sol earned 43.8% at the same named setting. The study covers seven models and includes five attempts on each task per evaluated setting. Reasoning effort and evaluation conditions matter when using those results to select a model.
Why experimental controls still matter
The report describes a case where Astra noticed that a control plate carried the wrong antibiotic label. In another case, Astra grouped sample bands correctly but selected the wrong colonies after assuming that a larger product meant success. The examples expose a difference between recognizing what an image shows and using the experiment’s controls to interpret it.
For teams evaluating an assistant, that difference suggests checking the reasoning behind a proposed measurement or selection. A useful review can ask the model to identify which control supports each conclusion, then have a scientist check that link before relying on the result.
