AI news ·
Medchembench evaluates AI models on medicinal chemistry reasoning tasks
GPT-6.2 Astra scored 84.5% on OpenAI's new 99-problem medicinal chemistry benchmark, while even the strongest models earned only about half the credit on compound selection and property prediction.

OpenAI released MedChemBench on November 20, 2026, a benchmark that tests how well AI models handle the core reasoning tasks of medicinal chemistry. The 99-problem set measures whether models can read molecular structures, interpret experimental results, and predict which compounds are worth pursuing next - a direct test of skills that could one day shorten the timeline for drug development.
GPT-6.2 Astra earned 84.5% of the available credit at the highest shared reasoning setting, compared with 83.3% for GPT-6 Astra, 82.0% for GPT-6.1 Sol, and 58.2% for GPT-5.6 Sol. The benchmark draws from 117 published papers and patents and includes 761 distinct molecular structures.
Why medicinal chemistry is hard for AI
Improving one property of a molecule often degrades another. A compound that blocks a disease protein strongly might be absorbed poorly or metabolized too fast. Chemists navigate these tradeoffs by designing, making, and testing related compounds, then deciding what to try next. For AI to assist, it must connect structural changes to experimental evidence and reason about which compounds deserve further work.
MedChemBench reflects this reality. Problems give models the biological target, the researchers' goals, compounds studied so far, and selected experimental findings. The model must then reconstruct molecular structures, explain results, rank compounds, and predict numerical properties.
Where Astra performed well
GPT-6 Astra showed its clearest strength in reconstructing molecular structures and explaining scientific evidence. In one spatial reasoning test, models had to assemble a molecule from a central structure and two attached groups while preserving an 18-atom ring. At the xhigh reasoning setting, GPT-6 Astra reconstructed the compound correctly in all ten responses. GPT-5.6 Sol produced a 15-atom ring, and GPT-5.6 Luna split the molecule into disconnected fragments.
"These errors change the identity of the molecule being analyzed," the researchers wrote. The task tests whether a model can accurately represent a compound before interpreting its behavior. Astra earned 95.4% of the available structure credit at xhigh.
Selection and prediction remain difficult
Even the strongest models struggled with compound selection and numerical property predictions. Astra earned roughly half the available credit in these categories. One problem, drawn from research on PRMT5 inhibitors studied in cancer drug discovery, asked models to predict how three related compounds would behave across multiple properties: target inhibition, cell barrier permeability, and metabolic breakdown rate.
Later experiments - withheld from the model - showed a tradeoff. Compound 23 blocked PRMT5 at a lower concentration, while compound 24 crossed the cell barrier more readily and broke down more slowly. Predicting these competing properties together tests whether a model connects molecular structure with several dimensions of compound behavior simultaneously.
What the benchmark does not measure
The results do not establish whether models can discover new drugs or improve laboratory decisions. Prior exposure to the published sources cannot be ruled out, and each response was a single attempt without tools, feedback, or new experiments. The benchmark evaluates reasoning from historical evidence rather than real-world drug discovery.
OpenAI evaluated eleven GPT models at different reasoning-effort settings. Higher reasoning settings improved scores, but large gaps between models persisted. GPT-6 Astra reached 83.8% at maximum effort. The component-level results reveal where models are advancing and where substantial room for improvement remains.
Why this matters for scientists and research professionals
MedChemBench isolates specific reasoning failures that matter in practice. A model that reconstructs a 15-atom ring instead of an 18-atom ring is analyzing the wrong molecule. A model that cannot weigh competing properties will miss the tradeoffs that define real drug discovery. For research scientists evaluating AI tools, benchmarks like this one provide a concrete way to assess whether a model can handle the structured reasoning their work demands - and where it still falls short. Professionals building these skills can explore structured learning paths through AI Scientific Research Courses or browse topic-specific offerings in AI for Scientists Courses.