OpenAI launches MedChemBench to test AI on medicinal chemistry

OpenAI on November 20 introduced MedChemBench, a benchmark that tests whether AI models can reason like medicinal chemists. Its top model, GPT-6.2 Astra, scored 84.5%, but only about half credit on picking candidate compounds and predicting their properties.
OpenAI on November 20 introduced a new benchmark called MedChemBench, designed to test whether AI models can reason like medicinal chemists — reading molecular structures, weighing experimental evidence to explain results, comparing compounds and predicting their properties. The results show models are strong at recognizing structures, but still weak at the step that actually drives drug research: choosing which compound to pursue.
99 problems, 761 molecular structures
MedChemBench contains 99 problems drawn from 117 published papers and patents, covering 761 distinct molecular structures. It evaluates four abilities: reconstructing molecular structures, scientific reasoning, ranking and selecting compounds, and predicting numerical properties. The problems span seven categories and focus on preclinical drug discovery, not clinical development.
Top model scores 84.5%, but only about half on compound selection
At the xhigh reasoning setting, GPT-6.2 Astra earned 84.5% of the available credit, ahead of GPT-6 Astra at 83.3% and GPT-6.1 Sol at 82.0%; the older GPT-5.6 Sol scored just 58.2%. The breakdown is telling: GPT-6 Astra earned 95.4% of the credit for reconstructing molecular structures, but only about 50.4% for ranking and selecting compounds and about 47.0% for numerical prediction — the model can draw the molecule accurately, but is far less reliable at judging which molecule is worth testing next.
OpenAI draws its own line
OpenAI states plainly in its release that the scores do not establish whether models can discover new drugs or improve laboratory decisions: the problems draw on published literature, so prior exposure during training cannot be ruled out, and each answer was a single attempt with no tools, feedback or new experiments. The company says MedChemBench is most valuable as a diagnostic tool for tracking where models are advancing and where they still fall short — not as proof that AI can replace drug discovery research.
It is OpenAI's second release in the same week measuring its models' professional abilities with a benchmark, after its November 18 post on using AI models to improve chip design — a sign that AI evaluation is moving from general question-answering toward specific expert domains.
N環球 
💬 Comments
No comments yet. Be the first!