OpenAI Tests Whether AI Agents Can Read Real Lab Data, and Its Best Model Scores 66.7%

The company built the tasks with Ligo, a biotech company focused on enzyme design. They span five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Agents are asked to describe what they see, check controls and avoid claims that go beyond the evidence. Physical lab work such as pipetting, operating instruments or running robots is outside its scope. Each task was reviewed by a second Ligo scientist and an external PhD-level wet-lab scientist, according to summaries of OpenAI's announcement by AlphaSignal and Kingy AI.

The benchmark tests a different skill from most biology evaluations, which ask models questions about biology rather than handing them messy experimental output. For drugmakers and research labs weighing AI agents for day-to-day analysis, that is closer to the work they would delegate. OpenAI also links these evidence-reading skills to both useful research and potential misuse, and recommends access controls, logging and review processes alongside accuracy testing, AlphaSignal reported.

According to the published results, OpenAI's GPT-6 Astra had the highest mean rubric score at 66.7%, with a 95% confidence interval of 60.3% to 73.1%. GPT-6.1 Sol followed at 62.0%, then GPT-5.6 Sol at 43.8%, 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%. Each model attempted every task five times. Scores include partial credit, so a 66.7% average does not mean two-thirds of tasks were fully solved.

The stronger models were also more concise. GPT-6 Astra produced 4,704 output tokens per task on average, against 27,620 for GPT-5.2, though token counts do not by themselves show cost or speed.

OpenAI illustrated the results with selected examples. GPT-6 Astra caught a plate labelled with a different antibiotic from the one used in the experiment, an error that would invalidate a control, and recognised that overlapping detection channels left a conclusion about cell death unsupported. Failures included a colony count about 12% too high on one plate, because the model missed colonies, merged neighbours and marked empty areas, and a case where a model correctly saw gel band patterns but assigned them to the wrong samples and so picked the wrong colonies.

The benchmark has clear limits. Only OpenAI models were tested, and the material reviewed by Kingy AI contained no measured human baseline. Kingy AI also found that the announcement did not link to a paper, dataset, full scoring rubrics, evaluation code or raw results on the day it was published, so outsiders cannot yet reproduce the scores. Whether OpenAI or Ligo will release the tasks for other labs to run is the main open question.