OpenAI Releases MedChemBench, Reports GPT-6.2 Astra at 84.5%
Summary
OpenAI releases MedChemBench, a 99-problem benchmark built from 117 papers and patents and 761 molecular structures that tests whether AI models can reconstruct molecules and predict compound properties. OpenAI reports GPT-6.2 Astra earns 84.5% of available credit at the xhigh setting, versus 58.2% for GPT-5.6 Sol. OpenAI says the results do not establish whether models can discover new drugs.
Key Points
- OpenAI reports GPT-6 Astra reaches 83.8% of available credit at the max reasoning setting.
- GPT-5.6 Sol draws a 15-atom ring instead of 18 on one structure task, OpenAI reports.
- OpenAI says prior exposure to the benchmark's published sources cannot be ruled out.