Science & Tech · AI · 9 hrs ago
OpenAI benchmark finds models struggle to predict drug compounds’ behavior
OpenAI introduced MedChemBench, a test of how well AI models handle complex medicinal chemistry tasks.
It uses research from 117 papers and patents to assess models on 99 problems involving 761 molecular structures.
The test found that GPT-6 Astra reconstructed structures with 95.4% of the available credit at its highest reasoning setting, but earned only about half the credit for choosing compounds and predicting numerical properties.
Knowing a molecule’s structure is not enough: drug researchers also need to understand how changes affect a compound’s activity, absorption and stability.
One test based on cancer research showed that related compounds could have different advantages, such as stronger activity against a protein or better passage across a cell barrier.
The results point to a gap in AI’s drug-research skills, but do not show whether models can discover drugs or improve laboratory decisions.
MedChemBench can help track progress, though its tasks use historical research and models worked without tools, feedback or new experiments.
OpenAI introduced MedChemBench, a benchmark for testing models’ reasoning on complex medicinal chemistry tasks.
The benchmark draws on 117 published papers and patents and covers 761 molecular structures across 99 problems.
At its highest reasoning setting, GPT-6 Astra earned 95.4 percent of available credit for reconstructing molecular structures.
The model earned only about half the available credit for selecting compounds and predicting their numerical properties.
The researchers said the benchmark identifies strengths and weaknesses but does not show that models can discover drugs or improve laboratory decisions.
- Who
- OpenAI introduced the benchmark; 11 models were evaluated.
- What
- MedChemBench tests models on medicinal chemistry tasks, including structure reconstruction, compound selection and property prediction.
- When
- Published November 21, 2026.
- Where
- Not stated.
- Why
- To assess models’ ability to reason about drug compounds and identify areas where they remain weak.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Benchmark
- MedChemBench
- Source material
- 117 published papers and patents
- Molecular structures
- 761
- Problems
- 99
- Models evaluated
- 11
- GPT-6 Astra structure reconstruction score
- 95.4 percent of available credit