News
November 22, 2026 · OpenAI

OpenAI's MedChemBench tests how well AI models reason about drug candidates

OpenAI's new MedChemBench grades models on 99 medicinal chemistry problems. GPT-6.2 Astra leads its tests but struggles with picking compounds and predicting properties.

OpenAI introduced MedChemBench on November 20, a benchmark that tests whether AI models can reason through medicinal chemistry problems. According to OpenAI's announcement, the benchmark checks whether a model can make sense of chemical structures and lab data and then decide which molecules deserve the next experiment. The problems come from published studies.

What's new

  • What it contains: OpenAI says MedChemBench has 99 problems built from 117 papers and patents, covering 761 distinct molecular structures.
  • Four skills graded: models are scored on rebuilding molecular structures correctly, explaining results from the evidence given, ranking and selecting compounds, and predicting numerical properties, with partial credit for predictions close to measured values.
  • Headline scores: at the xhigh reasoning setting, OpenAI reports that GPT-6.2 Astra earned 84.5% of available credit, against 83.3% for GPT-6 Astra, 82.0% for GPT-6.1 Sol and 58.2% for GPT-5.6 Sol.
  • Strengths and gaps: in OpenAI's breakdown, Astra scored 95.4% of the structure credit at xhigh but collected only around half the available points when it had to pick compounds or estimate measured values.
  • Effort helps: OpenAI tested eleven GPT models and says giving them more reasoning effort raised their scores, though wide differences between models persisted.
  • Stated limits: OpenAI cautions that the scores do not prove a model can find new medicines. It adds that the models may have seen the underlying papers during training, and that every answer was one try with no tools, no feedback and no fresh lab work.

Why it matters

Analysis: benchmarks like this help scientists judge where an assistant such as ChatGPT might be useful in a research workflow and where it still needs close checking. OpenAI's own numbers suggest models are already reliable at reading and describing chemistry, but much weaker at the forward-looking decisions that matter most in drug discovery. It is also a vendor-built benchmark that so far reports only OpenAI models, so independent results on other systems would make the comparison more useful.

Source: OpenAI