MedChemBench is here: on November 20, OpenAI released a medicinal-chemistry reasoning benchmark that tests whether models can think like medicinal chemists, using 99 problems drawn from real papers and patents. GPT-6.2 Astra scored 84.5% overall — but the headline number is not the most interesting part. The model is near-perfect at drawing molecular structures, yet its score drops to about half on the tasks that actually decide which drug candidate to pursue.
The benchmark's problems come from 117 papers and patents and cover 761 distinct molecular structures across seven categories, focused on preclinical drug discovery rather than clinical development. Scoring covers four areas: reconstructing molecules from structural drawings and chemical groups, explaining results from experimental evidence, ranking compounds and selecting candidates, and predicting numerical values and categories. So far OpenAI has tested only its own models: at the highest reasoning setting, xhigh, GPT-6.2 Astra scored 84.5%, GPT-6 Astra 83.3%, GPT-6.1 Sol 82.0%, and GPT-5.6 Sol 58.2%.
Near-Perfect on Structures, Only Half on Candidate Selection
The breakdown makes the weak spot clear. GPT-6 Astra earned 95.4% of the available credit for reconstructing molecular structures and 91.05% for scientific reasoning, but only 50.42% for ranking and selecting compounds and 47.02% for numerical predictions. In other words, models are good at drawing known molecules correctly and explaining the reasoning, but only halfway there when they must pick the most promising candidate from a list and attach reliable numbers to it — and that is exactly where medicinal chemists spend most of their time.
That matters for two audiences. Drug-discovery teams can use MedChemBench as a yardstick: lean on the model's strengths in structure and explanation, and keep human review in place for selection and prediction. General readers should watch out for one misreading — 84.5% does not mean AI can discover new drugs. OpenAI draws the line itself in the release: the results do not establish whether models can discover new drugs or improve laboratory decisions, prior exposure to the source papers cannot be ruled out, and each problem was answered in a single attempt, with no tools, feedback, or new experiments. Only OpenAI's own models have been scored so far, so there is no external comparison yet.
Two things are worth watching next: whether OpenAI opens the benchmark to third-party models, and whether it evolves toward open-ended tasks that mirror real research workflows. The exam room for medicinal chemistry is ready; the models are not yet on the job. For the model behind the top score, see our earlier breakdown of the GPT-6.2 Astra and Sol release.