What changed
OpenAI released MedChemBench with 99 problems drawn from 117 papers and patents and 761 molecular structures. It evaluates structure reconstruction, scientific reasoning, ranking and numerical prediction, and OpenAI reports GPT-6.2 Astra at 84.5% under its tested setup.
Why it matters
The benchmark creates a public, domain-specific test for whether models can connect molecular structure to experimental evidence instead of relying on generic science scores.
What to do
Download the benchmark and reproduce a subset with blinded inputs, contamination checks and domain-expert review. Track performance separately for reconstruction, explanation, ranking and numerical prediction before using results in research workflows.
Caveat
The 84.5% result is a vendor evaluation, not proof of drug-discovery capability. OpenAI says contamination cannot be ruled out; tests used single attempts without tools, feedback or experiments, and prediction and selection remain weaker.