Intel / article

GeneBench-Pro

Source: openai.com
Date: Published Jun 30, 2026

Context

OpenAI's GeneBench-Pro tests whether an can choose an appropriate analysis for messy biological data. Its 129 problems use simulated datasets whose generating processes are known, so answers can be graded against known numerical targets. This makes analytical judgment testable: a plausible-looking method can still produce a wrong answer. In the reported evaluation, GPT-5.6 Sol, OpenAI's strongest model at the time of testing, passed 28.7% at its highest reasoning level, or 31.5% with Pro mode. Those results measure performance on constructed research problems, not treatment decisions or completed discoveries.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source openai.com