OpenAI Introduces MedChemBench for Medicinal-Chemistry Reasoning

Unite.AI
Read full story· 5 min
OpenAI launched MedChemBench, a 99-problem benchmark for medicinal chemistry reasoning, where GPT-6.2 Astra scored 84.5% at the highest difficulty. The benchmark evaluates models on tasks like compound comparison and property prediction using real experimental data from drug discovery.

What happened

On November 20, 2026, OpenAI released MedChemBench, a benchmark consisting of 99 problems designed to test AI models on medicinal chemistry reasoning. The benchmark evaluates how well models can connect molecular structures with experimental results, compare different compounds, and predict properties using real data from drug discovery. OpenAI reported that its GPT-6.2 Astra model achieved an 84.5% score at the highest difficulty setting. The tasks are based on historical research and do not involve new experiments or real-time lab decisions.

Key details

  • MedChemBench includes 99 problems drawn from 117 published papers and patents.
  • GPT-6.2 Astra scored 84.5% at the xhigh reasoning-effort setting.
  • Other model scores at xhigh: GPT-6 Astra (83.3%), GPT-6.1 Sol (82.0%), GPT-5.6 Sol (58.2%).
  • The benchmark covers seven categories, including lead selection, structure–activity relationship forecasting, and ADME.
  • Grading weights: molecular structure (37.49%), scientific rationale (35.78%), ranking/candidate selection (17.51%), numerical/categorical prediction (8.92%).

Why it matters

Researchers and AI developers now have a standardized way to measure how well AI models can reason about medicinal chemistry problems using real-world data. Pharmaceutical companies and drug discovery teams can use these results to compare model performance on complex tasks like compound selection and property prediction. The benchmark reveals that current models are strong at reconstructing and explaining known problems but less reliable at selecting compounds or making quantitative forecasts, which may affect their immediate use in drug research workflows.


More in Healthcare