OpenAI’s MedChemBench tests AI on real drug-discovery decisions — and the scores split down the middle
OpenAI’s new benchmark grades AI models on the actual work of medicinal chemists: rebuilding molecules from drawings, explaining experimental evidence, and choosing which compounds to test next. The best model scores 84.5% overall — but only about half the credit on the decisions that matter most.
- MedChemBench: 99 problems drawn from 117 papers and patents, covering 761 distinct molecular structures.
- At the xhigh reasoning setting, GPT-6.2 Astra earned 84.5% of available credit, ahead of GPT-6 Astra at 83.3%, GPT-6.1 Sol at 82.0%, GPT-6.1 Terra at 75.1%, and GPT-5.6 Sol at 58.2%.
- GPT-6 Astra earned 95.4% of the available credit for reconstructing molecular structures, but only about half for selecting compounds and predicting numerical properties.
- The 18-atom ring test: GPT-6 Astra rebuilt the molecule correctly in all ten attempts; GPT-5.6 Sol made a 15-atom ring, and GPT-5.6 Luna split it into two fragments.
- Eleven GPT models were tested; each answer was a single attempt with no tools, feedback, or new experiments, and prior exposure to the published sources cannot be ruled out.
- OpenAI’s own caveat: the results do not establish whether models can discover new drugs or improve laboratory decisions.

Drug discovery runs on judgment calls. A medicinal chemist looks at a molecule, weighs what the last experiment actually showed, and decides which compound is worth making next — knowing that improving one property often breaks another. OpenAI’s newest benchmark tries to grade exactly that work. MedChemBench, introduced on November 20, hands models 99 problems drawn from real published research and scores them on the decisions chemists make between the first promising compound and a candidate worth testing in people.
The headline number looks strong. GPT-6.2 Astra, OpenAI’s newest model, earned 84.5% of the available credit. But the score is a composite, and the parts underneath it point in different directions: the models are near-flawless at reading and rebuilding molecules, and roughly a coin flip at the choices that decide what happens next. That split — not the ranking — is the news.
What the benchmark actually asks#
MedChemBench is built from 117 papers and patents and covers 761 distinct molecular structures. For most problems, a model is given what a chemist would have known at a chosen point in a research program: the biological target, the researchers’ goals, the compounds studied so far, and selected experimental findings. It is then graded on four separate abilities: reconstructing molecular structures from drawings and chemical fragments, explaining the science behind a result, ranking and selecting compounds, and predicting numerical properties such as potency and permeability.
One problem, drawn from research into PRMT5 inhibitors — a protein studied in cancer drug discovery — shows why the four parts matter together. The model must predict how three related compounds behave at once: how strongly each blocks PRMT5 inside cells, how readily each crosses a cell barrier, and how quickly liver enzymes break each one down. Later experiments, withheld from the model, showed the tradeoff in full: one compound blocked the target at a lower concentration, while another crossed the barrier more easily and broke down more slowly.

The scoreboard, and the split inside it#
OpenAI ran eleven GPT models through the benchmark at several reasoning-effort settings. The announcement reports scores at xhigh, the highest setting shared across the tested models:
| Model | Credit at xhigh |
|---|---|
| GPT-6.2 Astra | 84.5% |
| GPT-6 Astra | 83.3% |
| GPT-6.1 Sol | 82.0% |
| GPT-6.1 Terra | 75.1% |
| GPT-5.6 Sol | 58.2% |
At maximum effort, GPT-6 Astra reached 83.8%. But the component scores tell a sharper story. GPT-6 Astra earned 95.4% of the available credit for reconstructing molecular structures, and about 91% for explaining the scientific evidence. For ranking and selecting compounds it managed only about half the credit, and roughly the same for predicting numerical properties. OpenAI cautions that the categories use different tasks and scoring rules, so the percentages are not directly comparable as measures of difficulty — but the gap is the point. Reading a molecule is largely solved. Deciding what to do with it is not.
The 18-atom ring test#
Before a model can reason about a compound, it has to identify the molecule correctly — and chemical drawings pack a lot into a sketch: atom connections, ring structures, stereochemistry. One MedChemBench problem asks models to assemble a molecule from a central structure and two attached groups while preserving an 18-atom ring. At xhigh, GPT-6 Astra reconstructed the compound correctly in all ten responses. GPT-5.6 Sol produced a 15-atom ring instead; GPT-5.6 Luna split the molecule into two disconnected fragments.
The failure mode matters more than the score. A wrong ring is not a small error — it changes which molecule is being analyzed, so every conclusion downstream starts from the wrong compound. Getting the structure right is the foundation the rest of the reasoning stands on, and it is also where the newest models have pulled furthest ahead of the previous generation.

Why the caveats matter more than the ranking#
OpenAI is unusually direct about what MedChemBench does not show. The benchmark is retrospective: every problem comes from published work, so prior exposure to the source papers and patents cannot be ruled out — a model may have seen the answer’s neighborhood during training. Each response was a single attempt, with no tools, no feedback, and no new experiments. And only OpenAI’s own models were tested, so there is no independent comparison against other labs’ systems.
The company’s own sentence is the one to keep: “These results do not establish whether models can discover new drugs or improve laboratory decisions.” The paper behind the benchmark goes further, describing MedChemBench as close to saturation in its current form — the newest models are already near the ceiling on parts of it. What the benchmark does well is diagnose where models are strong and where they are weak. What it cannot do is certify a drug-discovery copilot.
What to watch#
Three things matter next. First, whether other labs and independent researchers reproduce the results on their own models — a benchmark run by one lab on its own systems is a starting point, not a leaderboard. Second, whether the next version of the test moves beyond retrospective problems toward prospective ones, where a model’s recommendation is recorded before the experiment runs. Third, whether the selection and prediction scores — the ones stuck near half credit — actually move as reasoning improves, or whether they expose a harder limit in how models weigh tradeoffs they have never run in a lab. Until then, MedChemBench is best read as a map of the gap, not a claim that the gap is closed.
Sources: OpenAI, 'Can AI help scientists identify promising drug candidates?' November 20, 2026; Unite.ai, 'OpenAI Introduces MedChemBench for Medicinal-Chemistry Reasoning,' November 20, 2026; Daily AI Brief, 'OpenAI launches MedChemBench to test AI models on medicinal chemistry reasoning,' November 20, 2026.