OpenAI released MedChemBench to test models on reading molecules and choosing what to test next.
Medicinal chemistry, the design of druglike molecules around competing properties, fails when potency improves but absorption or stability worsens.
The bench uses 99 problems from 117 papers and patents with 761 structures to grade structure, explanation, ranking and prediction per OpenAI.
At xhigh reasoning GPT-6.2 Astra earned 84.5% credit, ahead of GPT-6 Astra at 83.3% and GPT-5.6 Sol at 58.2% per OpenAI.
One task on PRMT5, a protein studied in cancer drug work, shows compound 23 hit harder while 24 crossed faster and degraded slower per OpenAI.
Astra kept 95.4% structure credit at xhigh but about half for selection and prediction without tools per OpenAI, so watch prediction not parsing.





