OpenAI Introduces GeneBench-Pro Benchmark for AI Agents in Computational Biology

OpenAI has introduced GeneBench-Pro, a research-level benchmark designed to evaluate how AI agents navigate ambiguity and make consequential, iterative judgments in computational biology, according to the OpenAI Blog. The suite expands on the earlier GeneBench with 129 problems across 10 domains—including statistical genetics, regulatory omics, clinical pharmacogenomics, and cancer genomics—to test what the authors term "research taste": the chains of judgment calls that shape a realistic scientific analysis.

To avoid pitfalls where benchmarks reward arbitrary choices or allow incorrect analyses to pass, each problem is synthetically generated from a known causal structure. This enables deterministic grading against known targets and ensures that plausible but incorrect analytic pathways fail. External reviewers, including faculty and graduate students from UCLA, assessed 82 of the 129 problems for realism and methodological soundness.

OpenAI reports that its strongest tested model, GPT-5.6 Sol, achieved a 28.7 percent pass rate at the highest reasoning level (31.5 percent with Pro mode enabled), up from below 5 percent for its best frontier model when the original GeneBench was first evaluated. The organization is open-sourcing 10 representative questions on Hugging Face and will provide a 50-question subset to Artificial Analysis for independent, third-party benchmarking.