OpenAI releases MedChemBench, a 99-problem medicinal chemistry test; GPT-6.2 Astra scores 84.5%

Built from 117 papers and patents with 761 molecular structures, the benchmark asks models to reconstruct molecules, explain results, rank compounds and predict properties; at the xhigh setting GPT-6.2 Astra earned 84.5% of credit against 58.2% for GPT-5.6 Sol, and only OpenAI models have been reported.
Why it matters
Medicinal chemistry is a core bottleneck in drug discovery because improving one property of a compound can worsen another. MedChemBench tries to measure whether models can reason through those trade-offs using evidence from real research programmes, such as PRMT5 inhibitors studied in cancer. The scores show steep gains from GPT-5.6 Sol to GPT-6 Astra but little gap among the newest models. Watch whether other labs report results, since only OpenAI models appear so far.
Key facts
- 99 problems from 117 papers and patents
- 761 distinct molecular structures
- GPT-6.2 Astra: 84.5% at xhigh
- GPT-6 Astra 83.3%, GPT-6.1 Sol 82.0%, GPT-5.6 Sol 58.2%