openai.com web signal

OpenAI releases MedChemBench; GPT-6.2 Astra leads at 84.5%

TL;DR

  • MedChemBench is 99 problems drawn from 117 papers and patents, covering 761 distinct molecular structures across four task types.
  • At xhigh reasoning, GPT-6.2 Astra scored 84.5% of available credit, with GPT-6 Astra at 83.3% and GPT-6.1 Sol at 82.0%.
  • GPT-6 Astra reconstructed molecular structures 95.4% of the time but earned roughly 50% on compound selection and numerical property prediction.

OpenAI's MedChemBench, published November 20, is 99 problems drawn from 117 papers and patents and covering 761 distinct molecular structures. It asks a model to read molecular structures, back explanations with supplied evidence, rank and pick compounds, and predict numerical properties.

At the xhigh reasoning setting, GPT-6.2 Astra scored 84.5% of available credit, just ahead of GPT-6 Astra at 83.3% and GPT-6.1 Sol at 82.0%. GPT-5.6 Sol trailed at 58.2%. The shape of the error says more than the ranking does: GPT-6 Astra reconstructed molecular structures correctly 95.4% of the time, but on compound selection and numerical property prediction it earned only around half credit.

OpenAI publishes the ceiling alongside the score. "These results do not establish whether models can discover new drugs or improve laboratory decisions," the post states. "Prior exposure to the published sources cannot be ruled out, and each response was a single attempt without tools, feedback, or new experiments."