MedChemBench at a glance
OpenAI introduced MedChemBench on November 20, 2026, a benchmark that tests whether AI can reason about medicinal chemistry. The top system, GPT-6.2 Astra, scored 84.5% on the benchmark, an OpenAI evaluation of drug discovery reasoning.
What changed with MedChemBench
OpenAI announced on November 20, 2026, in its introduction, that MedChemBench is a new public benchmark of more than 800 medicinal chemistry problems and more than 4,800 model solutions, built by the drug discovery company Isomorphic Labs with Oxford University, King's College London and AstraZeneca. The problems are designed to capture the reasoning a medicinal chemist uses, and each one is tested against a real molecule in an experimental assay that measures the activity or toxicity of a compound.
MedChemBench measures an AI on 10 tasks, from proposing new molecules to synthesising them, with 250 compounds already tested in the lab. OpenAI said GPT-6.2 Astra at its highest reasoning level scores 84.5% on the benchmark, while the GPT-6 Astra model scores 83.3% at the same level. It also said MedChemBench is a test of AI drug discovery reasoning and that OpenAI plans to build on the results by giving its science models extra chemical information, such as the 3D shape of molecules, and testing them on other scientific subjects.
MedChemBench in numbers
Values are copied from the linked source on the date shown. They are not IA Compa measurements.
- MedChemBench score, GPT-6.2 Astra at xhigh (OpenAI's evaluation) — 84.5% (as of Nov 20, 2026)
- MedChemBench score, GPT-6 Astra at xhigh (OpenAI's evaluation) — 83.3% (as of Nov 20, 2026)
Why MedChemBench matters for businesses
Why it matters: businesses can now check an AI's work against a public benchmark before giving it research that touches real molecules. The 84.5% result shows the strongest model still fails about one in six tasks, so a company using AI to propose new molecules will still need a chemist to check each proposal before it enters the lab.

