OpenAI Introduces MedChemBench for AI Chemistry Testing
OpenAI’s MedChemBench tests how AI models reconstruct molecular structures, use experimental evidence and evaluate compounds in medicinal chemistry. The company introduced the benchmark publicly on November 20, 2026, as an assessment of reasoning on medicinal-chemistry problems.
The evaluation is retrospective, according to OpenAI’s technical paper dated November 9, 2026. It draws on 117 papers and patents to assemble 99 problems involving 761 distinct molecular structures. OpenAI states that the results do not establish whether models can discover new drugs or improve decisions in a laboratory.
Most problems present a drug-development program at a particular stage. Models receive its biological target and research objectives, alongside compounds already studied and selected experimental data. According to the paper, findings reported later in the underlying publication are excluded from the prompt but used to prepare reference answers and determine scores.
The assessment covers reasoning from evidence, reconstructing structures, ranking or choosing compounds, and predicting molecular properties. Its seven categories span molecular design, safety, lead selection, assay biology, retrosynthesis, structure-activity forecasting, and ADME and delivery. The paper describes medicinal chemistry as requiring trade-offs among potency, selectivity, exposure, safety and how readily a compound can be synthesized: changing a structure can improve one property while worsening another.
In its announcement, OpenAI reported that GPT-6.2 Astra received 84.5% of the available credit with reasoning set to xhigh. The comparison included GPT-6.1 Sol at 82.0%, GPT-5.6 Sol at 58.2% and GPT-6 Astra at 83.3%.
Performance differed by task. For GPT-6 Astra at xhigh, OpenAI reported a structure-reconstruction score equal to 95.4% of available credit. Compound selection and numerical-property prediction each yielded roughly half the credit available for those tasks.
The technical paper presents a separate set of results. Its highest reported mean was 83.83%, recorded by GPT-6 Astra using max effort. All eleven GPT models evaluated in the paper were also tested at xhigh, where their scores extended from 35.02% to 83.33%.
OpenAI’s description places limits on what those scores demonstrate. Models had one attempt at each response, with neither tools nor feedback available and no new experiments conducted. The authors also stated that they could not rule out exposure during training to the published material used to construct the benchmark.
* This article is based on publicly available information at the time of writing.