OpenAI Publishes BioBasicsBench for AI Lab Evidence Tests
BioBasicsBench tests AI agents on interpreting experimental data, identifying problems and determining when a conclusion lacks sufficient evidence, according to OpenAI. The company published the 100-task benchmark on November 6, 2026, and said it developed the assessment with Ligo, a biotechnology company focused on enzyme design.
OpenAI reported that GPT-6 Astra earned an average of 66.7% of available credit at the xhigh reasoning-effort setting, compared with 43.8% for GPT-5.6 Sol. Those figures measure credit earned under the grading criteria, rather than the proportion of tasks solved completely. The company said better benchmark scores alone do not establish that a model can make dependable laboratory judgments.
Under OpenAI’s scoring method, an answer can receive separate credit for its measurements, conclusion, reasoning and recognition of insufficient evidence. The reported mean reward is the average task score. To calculate 95% confidence intervals, the company resampled tasks as complete units, keeping the five attempts associated with each task together.
Two examples in OpenAI’s account describe the models’ handling of experimental evidence. For a cloning exercise, GPT-6 Astra identified a mismatch between a plate marked kanamycin and an experiment using ampicillin, and rejected the plate as a dependable control. GPT-5.6 Sol instead used the lack of colonies on that plate in its assessment.
A separate colony-counting exercise produced errors from GPT-6 Astra, according to OpenAI. Its total exceeded the benchmark’s reference count by about 12%. The model omitted some visible colonies, did not distinguish some neighboring colonies from one another and included areas of empty agar in its count.
The benchmark covers molecular cloning and nucleic acid methods, alongside protein methods, imaging and cytometry, and quantification and assays. OpenAI said tasks can require several kinds of evidence together. Materials include sequencing data, gel images, microscopy, colony plates and outputs from laboratory instruments.
Across the evaluation, OpenAI reported 18,500 results from 37 combinations of model and reasoning effort, with five attempts per task for each combination. Seven models were tested: GPT-6 Astra, GPT-6 Luna, GPT-6.1 Sol, GPT-5.6 Sol, GPT-5.5, GPT-5.4 and GPT-5.2.
According to OpenAI, scientists used results from their own experiments to construct the exercises, with experts supplying the grading criteria for each. The review process included a Ligo scientist other than the task’s author and an external scientist with a PhD and wet-lab expertise, followed by a final internal review.
* This article is based on publicly available information at the time of writing.