OpenAI has introduced MedChemBench, an internal benchmark designed to measure how well AI models perform medicinal chemistry tasks. Its tests cover chemical structure interpretation, predictions of compound potency and toxicity, candidate selection, and synthesis route planning. As of late November 2026, GPT-6 Astra leads the leaderboard with 49.7%, followed by GPT-5.6 Sol at 48.3%, with no model reaching the halfway mark. This ceiling highlights persistent difficulties in consistent multi-step chemical reasoning. In June 2026, OpenAI also deployed GPT-5.4 with Molecule.one to run 10,080 reactions aimed at improving yields in a Chan-Lam coupling.
Source: Read the original article


