OpenAI introduced MedChemBench on November 20, 2026, a benchmark that tests whether AI models can reason like medicinal chemists. The task set is built from published research: 99 problems drawn from 117 papers and patents, covering 761 distinct molecular structures. Each problem gives a model the context available at one point in a drug-discovery project, such as the biological target, the researchers' goals, the compounds studied so far and selected experimental findings.
Medicinal chemistry is hard to automate because improving one property of a molecule often damages another. A compound can act strongly on a disease-related protein yet be absorbed poorly or broken down too fast in the body. OpenAI says useful AI help therefore has to link changes in structure to experimental evidence and judge which compounds are worth making next.
The benchmark grades four abilities: rebuilding molecular structures with the right atom connections and stereochemistry, giving scientific explanations supported by the evidence supplied, ranking and selecting compounds, and predicting numerical properties against values reported in the source material. Predictions earn partial credit according to how close they are to the measured result. One example concerns inhibitors of PRMT5, a protein studied in cancer research, where the model must predict how three related compounds will behave across three properties and a later experiment showed a trade-off between potency and permeability.
On the headline result, OpenAI reports that GPT-6.2 Astra earned 84.5 percent of the available credit at its xhigh reasoning setting, against 83.3 percent for GPT-6 Astra, 82.0 percent for GPT-6.1 Sol and 58.2 percent for GPT-5.6 Sol. Eleven GPT models were evaluated at several reasoning settings, and higher settings improved scores while large gaps between models remained. In one structure-building problem that required keeping an 18-atom ring, GPT-6 Astra succeeded in all ten responses, while GPT-5.6 Sol produced a 15-atom ring and GPT-5.6 Luna split the molecule into two fragments.
The detailed results show where the weak spots are. Astra earned 95.4 percent of the structure credit at xhigh, and OpenAI says it was strongest at reconstructing molecules and explaining experimental evidence. For selecting compounds and predicting numerical properties it earned only about half of the available credit, which OpenAI describes as substantial room to improve.
OpenAI is careful about what the numbers do not show. The results do not establish that models can discover new drugs or improve decisions in a laboratory. Prior exposure to the published sources cannot be ruled out, and each response was a single attempt without tools, feedback or new experiments. All figures come from OpenAI's own paper and evaluation, and no independent replication has been reported.