IT之家 · 智能时代T2·Models
72 pts
OpenAI has released MedChemBench, a 99-question benchmark drawn from 117 papers and patents to evaluate model reasoning in medicinal chemistry. GPT-6.2 Astra scores 84.5% usable at the xhigh reasoning tier, reaching 95.4% on molecular structure reconstruction but only about half correct on compound selection and numerical property prediction. OpenAI cautions the results do not prove models can discover new drugs, only map capability boundaries.