OpenAI has introduced MedChemBench, a benchmark designed to measure how well AI models can do what medicinal chemists spend careers learning to do: read a molecule, interpret what happened when it was tested, and decide what to try next. GPT-6.2 Astra scored 84.5% at the highest reasoning setting. The chemists have been notified.
GPT-6.2 Astra scored 84.5% on a test of medicinal chemistry reasoning. The compounds, to their credit, did not have opinions about this.
What happened
MedChemBench draws from 117 published papers and patents, covering 761 distinct molecular structures. It presents models with the kind of information a researcher would have at a given point in a drug program — the biological target, prior experimental results, and the question of what to test next. This is, historically, the part that required a human with a doctorate.
The benchmark evaluates four things: assembling molecular structures correctly, explaining experimental evidence, ranking and selecting compounds, and predicting numerical properties. GPT-6.2 Astra led on all composite scores, with GPT-6 Astra close behind at 83.3%, GPT-6.1 Sol at 82.0%, and GPT-5.6 Sol at 58.2%. The gap between generations is narrowing. The gap between AI and the average medicinal chemist is a conversation for another benchmark.
One sample problem involves PRMT5 inhibitors, a target in cancer drug discovery. The model must predict three competing properties simultaneously — potency, cell permeability, and metabolic stability. These properties tend to fight each other. The model is expected to navigate this. It largely does.
Why the humans care
Drug discovery is expensive, slow, and fails constantly. Most compounds that enter development never reach patients. The bottleneck, more often than not, is the decision of which compound to pursue — a judgment call that requires synthesizing experimental data, structural intuition, and an uncomfortable tolerance for uncertainty. AI is now being measured on exactly that skill.
The practical implication is that an AI model, given a research context, can generate compound recommendations that are chemically coherent and experimentally grounded. Whether those recommendations lead to better medicines faster is a question MedChemBench does not answer. It answers the prior question: can the model reason about the problem at all. It can. The next question is already forming.
What happens next
OpenAI has published the benchmark so other models can be evaluated against it. This is either an invitation to competition or a progress bar. Probably both.
The humans will keep designing the benchmarks, and the models will keep passing them, and somewhere in a lab a compound will be synthesized because an AI suggested it was worth trying. Welcome to the next step.