OpenAI introduces BioBasicsBench for wet-lab AI agents

OpenAI introduced BioBasicsBench, a benchmark measuring how well AI agents interpret raw experimental data across 100 routine wet-lab tasks. Developed with the biotech company Ligo, the evaluation tests whether agents can interpret evidence, identify errors, and recognize when observations cannot support a conclusion. The benchmark covers five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning.

The tasks use raw experimental evidence, including 31 tasks for gels and blots, 27 for protocols and notes, and 20 for tables and instrument exports. Other data types include 14 microscopy tasks and 12 sequencing files. Scientists created the tasks using data from their own experiments, and each underwent review by a PhD-level scientist working in a wet lab.

OpenAI evaluated seven models across 37 model and reasoning-effort settings, producing 18,500 results. At the 'xhigh' reasoning setting, GPT-6 Astra earned an average of 66.7% of the available credit. This result compared with 62.0% for GPT-6.1 Sol and 43.8% for GPT-5.6 Sol. Other models tested at the same effort level included GPT-6 Luna at 38.3%, GPT-5.5 at 35.4%, GPT-5.4 at 33.2%, and GPT-5.2 at 31.2%.

The evaluation showed that GPT-6 Astra identified the limits of what specific data shows. In one instance, Astra recognized that brightness in cell images was ambiguous due to overlapping detection windows between 525–540 nm. It also correctly identified a z-stack in a series of microscope images, whereas GPT-5.6 Sol looked for a biological explanation for cells appearing to disappear.

The report noted that higher scores do not yet mean models can make reliable judgments in the lab. In some cases, assumptions can override evidence. For example, Astra selected lanes 4, 6, 7, and 8 in a PCR product test based on the assumption that a larger product identifies success, while the reference selected lanes 5, 9, and 10.