Summary
On June 30, 2026 OpenAI introduced GeneBench-Pro, a research-level benchmark of 129 synthetic genomics problems across 10 domains that tests whether AI agents can diagnose messy biological data, pick the right analysis path, and reach actionable conclusions; the top model GPT-5.6 Sol Pro passed only 31.5 percent at maximum reasoning.
What changed
OpenAI published GeneBench-Pro, a benchmark of 129 synthetically generated computational-biology problems with known causal ground truth across 10 domains, measuring agent judgment on measurement error, bias, confounding, QC failures, and model selection; GPT-5.6 Sol Pro scored 31.5 percent and GPT-5.6 Sol 28.7 percent at maximum reasoning.
Why it matters
It reframes AI-for-science progress around judgment under uncertainty rather than knowledge recall, and the low pass rates set a public bar showing frontier agents remain far from autonomous research, useful context as vendors like Anthropic push agents into life sciences.
Evidence excerpt
a research-level benchmark for a harder kind of AI progress: how well agents can navigate messy biological data, choose the right analysis path, and make judgment calls that real computational research depends on.