New OpenAI benchmark tests how well models reason through complex drug discovery tasks
Advanced computing models can reconstruct complex molecular structures with high accuracy, but predicting how potential...

Advanced computing models can reconstruct complex molecular structures with high accuracy, but predicting how potential drugcompounds behave remains a challenge, according to a new medicinal chemistry benchmark introduced by OpenAI.
Called MedChemBench, the evaluation tests whether models can connect molecular structures with experimental evidence to explain research findings, compare compounds, and predict their properties.
The benchmark draws on 117 published papers and patents and includes 761 distinct molecular structures across 99 problems.
At the xhigh reasoning setting, GPT-6 Astra earned 95.4 percent of the available credit for reconstructing molecular structures. However, the model earned only about half the available credit for selecting compounds and predicting their numerical properties.
The findings highlight a gap between identifying chemical structures correctly and understanding how changes to those structures influence a potential drug’s behavior.
Testing complex chemical trade-offs
Medicinal chemists routinely balance competing properties when developing drug candidates. A compound may effectively block a disease-related protein but perform poorly when absorbed by the body or break down too quickly.
Improving one characteristic can also weaken another, making it difficult to identify compounds that meet several requirements simultaneously.
MedChemBench evaluates four capabilities: molecular structure reconstruction, scientific reasoning, compound ranking and selection, and numerical property prediction.
One of its problems draws on research into inhibitors of PRMT5, a protein studied in cancer drug discovery.
Models receive molecular structures and experimental results from earlier stages of the research. They must predict how three related compounds will perform in terms of PRMT5 inhibition in cells, permeability across a cell barrier, and breakdown by liver enzymes.
Later experimental results, withheld from the models, revealed competing advantages. Compound 23 inhibited PRMT5 at a lower concentration, while compound 24 crossed the cell barrier more readily and degraded more slowly in the laboratory assay.
The task tests whether models can recognize that stronger activity against a biological target does not necessarily translate into better absorption or metabolic stability.
Molecular accuracy remains crucial
The benchmark also examines whether models can reconstruct molecules accurately from chemical drawings, including their atom connections, ring structures and stereochemistry.
In one example, models were asked to assemble a compound while preserving an 18-atom ring.
GPT-6 Astra reconstructed the structure correctly in all 10 responses at the highest reasoning setting shared by the evaluated models. In contrast, GPT-5.6 Sol produced a smaller, 15-atom ring, while GPT-5.6 Luna generated disconnected molecular fragments.
Such errors can change the identity of the compound being analyzed, undermining subsequent predictions about its properties.
Despite its strong structural performance, the model struggled more with compound selection and numerical predictions. These tasks require connecting molecular changes with several experimental measurements rather than simply reproducing a chemical structure.
The researchers evaluated 11 models across different reasoning settings. Higher reasoning effort generally improved performance, although substantial differences remained between models.
The team cautioned that the results do not establish whether models can discover new drugs or improve laboratory decisions. The benchmark relies on historical research, and prior exposure to published sources cannot be ruled out. Models also completed the tasks without tools, feedback, or new experiments.
MedChemBench therefore offers a way to identify progress and remaining weaknesses in computational reasoning for medicinal chemistry, rather than demonstrating that these systems are ready to replace drug discovery research.
Get the latest news delivered to your inbox.
Related articles





