OpenAI MedChemBench tests AI reasoning for drug research
OpenAI's MedChemBench benchmark tests whether AI models can reconstruct molecules, interpret experiments and choose promising compounds for medicinal chemistry.
OpenAI has introduced MedChemBench, a benchmark designed to test whether AI models can reason about molecular structures, experimental evidence and the trade-offs involved in choosing compounds for further medicinal-chemistry work.1
What MedChemBench tests
The benchmark contains 99 problems drawn from published research, covering 117 papers and patents and 761 distinct molecular structures. OpenAI says the tasks ask models to reconstruct molecules, explain experimental results, rank or select compounds and predict reported properties.1
That combination is intended to test more than whether a model can recognise a chemical diagram. A model must connect structural changes to several competing outcomes, such as activity against a target, permeability through a cell barrier and how quickly liver enzymes break a compound down. Improving one property can make another worse, which is a routine problem in medicinal chemistry.1
Stronger at structures than predictions
OpenAI reports that GPT-6.2 Astra earned 84.5% of the available credit at the xhigh reasoning setting, ahead of GPT-6 Astra at 83.3%, GPT-6.1 Sol at 82.0% and GPT-5.6 Sol at 58.2%. Astra's structure score reached 95.4%, while compound selection and numerical property prediction remained much harder.1
One example asks a model to preserve an 18-atom ring while assembling a molecule from a central structure and two attached groups. OpenAI says GPT-6 Astra reconstructed it correctly in all ten responses at the shared xhigh setting, while GPT-5.6 Sol produced a 15-atom ring and GPT-5.6 Luna split the molecule into disconnected fragments.1
OpenAI cautions that the benchmark does not establish whether models can discover new drugs or improve laboratory decisions. The results use historical evidence, prior exposure to the source material cannot be ruled out, and the evaluations used single attempts without tools, feedback or new experiments.
Our opinion
MedChemBench is more useful as a map of where AI reasoning breaks than as proof that a model is ready to direct drug discovery. The gap between identifying structures and choosing compounds shows why scientific benchmarks need to test the full chain from representation to measured, real-world outcomes.