OpenAI released MedChemBench, a benchmark that measures how well AI models reason through medicinal chemistry problems, built from 99 problems drawn from published drug research. OpenAI's newest model, GPT-6.2 Astra, earned 84.5% of available credit overall at the xhigh reasoning setting, according to the company.
The problems draw on 117 papers and patents and include 761 distinct molecular structures. For most problems, MedChemBench gives a model the context available at a chosen point in a real research program: the biological target, the researchers' goals, the compounds studied so far and selected experimental findings.
The benchmark scores four capabilities. Structure reconstruction tests whether a model gets atom connections and stereochemistry right. Scientific reasoning checks whether explanations are supported by the supplied evidence. The other two cover ranking and selecting compounds, and predicting properties against reported numerical values and categories, with partial credit based on how close a prediction lands to the measured value.
Key facts
- 01MedChemBench has 99 problems built from 117 papers and patents, covering 761 distinct molecular structures.
- 02GPT-6.2 Astra earned 84.5% of available credit overall at the xhigh setting; GPT-5.6 Sol scored 58.2%.
- 03GPT-6 Astra scored 95.4% on molecular structure reconstruction but about half the credit on selection and property prediction.
- 04OpenAI evaluated eleven GPT models across reasoning-effort settings, with single attempts and no tools.
- 05OpenAI says the results do not establish whether models can discover new drugs or improve lab decisions.
OpenAI evaluated eleven GPT models at different reasoning-effort settings, with xhigh the highest setting all eleven shared. At xhigh, GPT-6 Astra scored 83.3% overall, GPT-6.1 Sol 82.0%, GPT-6.1 Terra 75.1% and GPT-5.6 Sol 58.2%. GPT-6 Astra reached 83.8% at max effort, and higher effort generally improved results.
Structure work was the strongest area. GPT-6 Astra earned 95.4% of available credit on structure reconstruction at xhigh. On one problem involving a PRMT5-targeting molecule with an 18-atom ring, GPT-6 Astra was correct in all ten xhigh responses, while GPT-5.6 Sol and GPT-5.6 Luna failed all ten. In examples OpenAI highlighted, Sol produced a 15-atom ring and Luna split the molecule into two disconnected fragments.
Compound selection and numerical property prediction were harder. GPT-6 Astra earned only about half the available credit on those components, which require connecting molecular changes with several experimental measurements at once. OpenAI cautioned that category percentages are not directly comparable because each category uses different tasks and scoring rules.
One property task asked models to predict cellular PRMT5 inhibition, cell-barrier permeability and liver-enzyme breakdown rate for three related compounds. Experimental data withheld from the models showed compound 23 blocked PRMT5 in cells at a lower concentration, while compound 24 crossed the cell barrier more readily and broke down more slowly.
OpenAI attached caveats to the results. The company said they do not establish whether models can discover new drugs or improve laboratory decisions. Because the benchmark is built on historical research, prior exposure to the published sources cannot be ruled out, and each response was a single attempt with no tools, feedback or new experiments.
OpenAI published a preprint with additional examples alongside the benchmark. In announcing MedChemBench, OpenAI said it shows the progress its frontier models have made in helping scientists use AI to identify promising drug candidates.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




