OpenAI tests whether AI can reason in biology

OpenAI is moving AI evaluation from school-like tasks toward work now performed by scientific teams. GeneBench-Pro contains 129 tasks across ten areas of computational biology, including statistical genetics, pharmacogenomics, clinical diagnostics, proteomics and cancer genomics. Each task gives the model data, context and a decision goal. The important point is that OpenAI is not only measuring whether the final number is correct. The benchmark targets judgment. In biology, a result often depends on whether a researcher detects noise, a technical error, an unsuitable method or weak evidence. That matters for pharma and biotech, where faster analysis can save money, but a wrong conclusion can undermine an entire decision. The results show rapid progress and clear limits. GPT-5.6 Sol reached 28.7 percent success in the highest reasoning mode and 31.5 percent in Pro mode. OpenAI says the best model reached less than five percent when the original GeneBench was created. Models are improving quickly, but they still fail on most tasks. The business impact is significant. Reviewers estimated that one task would take a human expert 20 to 40 hours. At 200 dollars per hour, that means thousands of dollars of work. OpenAI says inference costs only a few dollars per task. Even partly reliable automation can therefore reduce the cost of hypothesis screening, analysis review and research preparation. Regulation will still require an audit trail. A bad biological analysis can affect drug development, diagnostics and investment decisions. Companies therefore cannot treat such models as replacements for scientists. They are more likely to become tools that accelerate work while still requiring verification, responsibility and a clearly documented process. Terms: A benchmark is a test set for comparing model capabilities. Pharmacogenomics studies how genetics affects drug response. Inference is the operational use of a model for a specific task.

OpenAI tests whether AI can reason in biology

The full article is available to users registered at Hard Skills