GeneBench-Pro tests the part of research that usually disappears inside expert judgment: choosing the right analysis, spotting weak data, revising assumptions, and knowing when an answer is ready to support a decision.
OpenAI frames the benchmark around computational biology tasks where scientific data are messy and the path is not fully specified. That is a different bar from asking a model to recall facts or follow a clean workflow.
The benchmark moves closer to real research work
OpenAI says GeneBench-Pro covers 129 questions across genomics, quantitative biology, translational medicine, and related domains. Each problem gives the agent a dataset, context, and a target estimand tied to a downstream decision.
That setup forces the model to do more than run code. It has to decide what the data can support, which analysis path fits the question, whether diagnostics change the plan, and how to close the loop without hiding uncertainty.
Judgment becomes measurable
The useful shift is measurement. Scientific AI systems will not be trusted only because they finish tasks quickly. They need to show where their reasoning holds, where it breaks, and which failures are caused by weak assumptions rather than missing knowledge.
GeneBench-Pro turns that vague gap into something more inspectable. OpenAI reports that GPT-5.6 Sol reaches 28.7% at the highest reasoning level and 31.5% with Pro mode, while earlier frontier systems scored far lower on the original GeneBench.
Those numbers are not a replacement for expert review. They show how much hard research judgment remains. The practical value is in making that gap visible enough to improve.
Related services: Operations, Quality, OpenAI expertise