OpenAI and enzyme-design company Ligo have introduced BioBasicsBench, a benchmark of 100 routine wet-lab analysis tasks. Published on November 6, it asks AI agents to interpret experimental evidence, detect errors and recognize when the data cannot support a conclusion.
The announcement reports 66.7% of available rubric credit for GPT-6 Astra at the xhigh reasoning setting, compared with 43.8% for GPT-5.6 Sol. The percentage is mean reward with partial credit, rather than the share of tasks solved completely. The tasks cover raw material such as gels, colony plates, microscopy and instrument outputs.
OpenAI also shows failures, including incorrect experimental interpretation and colony counting. The results are the developer’s own evaluation of its models. The announcement does not provide a public task set or evaluation code with which this article could independently reproduce the scores.
Why controls matter as much as recognition
Our analysis is that the benchmark focuses attention on a demanding part of laboratory reasoning: deciding whether an observation is a valid basis for a conclusion. Reading an image correctly can still lead to the wrong decision when the control is inappropriate or the experiment’s context is misunderstood.
This is useful for teams building scientific assistants because it suggests more specific acceptance tests. An agent should identify the evidence it used, check the conditions required for that interpretation and acknowledge missing measurements. A plausible explanation without those links can waste time on the next experiment.
Partial-credit scoring is helpful for studying progress, but it requires careful reporting. A system may earn credit for noticing a relevant feature while still choosing the wrong final action. Product teams should inspect those combinations rather than treating the aggregate as a deployment-readiness score.
What a laboratory can test locally
A practical evaluation can use familiar completed experiments with expert-reviewed outcomes. Include failed runs and ambiguous observations, alongside clear cases. Measure whether the assistant reaches the supported conclusion and whether it appropriately requests additional evidence.
Keep analysis assistance separate from authorization to operate equipment. The ability to interpret a picture does not establish that an agent can safely execute a protocol or manage a physical workflow.
BioBasicsBench is a useful new target for measuring scientific assistance because it treats everyday evidence and uncertainty as central. Its reported gains show progress within the stated evaluation, while the failure cases explain why human review and reproducible local testing remain important.