AI
OpenAI's MedChemBench Tests If AI Can Help Pick Drug Candidates
OpenAI's MedChemBench tests whether AI can help scientists identify drug candidates, using 99 problems from published research.

Key Takeaways
- 1OpenAI launched MedChemBench on November 20, 2026, to evaluate AI in medicinal chemistry.
- 2MedChemBench uses 117 papers and 761 molecular structures for realistic AI evaluation.
- 3On one problem preserving an 18-atom ring, GPT-6 Astra was right in all ten responses; GPT-5.6 Sol and Luna failed.
- 4OpenAI says the results do not establish whether models can discover new drugs or improve lab decisions.
- 5UCLA researchers report their MAESTRO platform evaluates each drug candidate more than 100 times.
OpenAI introduced MedChemBench on November 20, 2026, a benchmark that tests how well AI models read molecular structures, interpret experimental results, and choose which compounds to test next. The question it asks is whether AI can help scientists identify promising drug candidates.
This new benchmark aims to test AI's ability to connect changes in molecular structure with experimental evidence. Medicinal chemists constantly navigate complex trade-offs, where improving one molecular property often degrades another, making compound selection difficult.
OpenAI calls building AI that can reliably support this work one step towards helping scientists develop better medicines. MedChemBench measures these abilities through 99 problems drawn from published research. At the xhigh reasoning setting, GPT-6.2 Astra earned 84.5% of the available credit, compared with 83.3% for GPT-6 Astra, 82.0% for GPT-6.1 Sol, and 58.2% for GPT-5.6 Sol.
OpenAI notes that these results do not establish whether models can discover new drugs or improve laboratory decisions. Prior exposure to the published sources cannot be ruled out, and each response was a single attempt without tools, feedback, or new experiments.
How Does MedChemBench Evaluate AI Models?
MedChemBench evaluates AI models on four core components: reconstructing molecular structures, providing scientific reasoning for experimental findings, ranking and selecting compounds based on stated requirements, and predicting numerical properties. This benchmark draws on 117 scientific papers and patents, featuring 761 distinct molecular structures.For most problems, models receive contextual information available at a chosen point in a research program. This includes the biological target, research goals, previously studied compounds, and selected experimental findings. The benchmark assesses how closely the model's predictions align with reported values and evidence.
| Evaluation Component | Description |
|---|---|
| Molecular Structures | Correct atom connections and specified stereochemistry from drawings and chemical groups. |
| Scientific Reasoning | Explanations supported by supplied evidence for compound choices or experimental results. |
| Ranking and Selection | Compound rankings and choices that meet specific research requirements. |
| Property Predictions | Agreement with numerical and categorical values reported for molecular properties. |
Testing spatial reasoning is fundamental; models must correctly identify molecules before interpreting their behavior. In one problem that required preserving an 18-atom ring, GPT-6 Astra reconstructed the compound correctly in all ten responses at the xhigh setting, while GPT-5.6 Sol and GPT-5.6 Luna failed across all ten. GPT-6 Astra earned 95.4% of the structure credit at xhigh, but only about half the credit for selecting compounds and predicting numerical properties.
How Others Are Applying AI to Drug Discovery
Outside the benchmark, Revvity Signals describes machine learning models that evaluate large chemical libraries and help predict which compounds are most likely to show desired biological activity before laboratory testing begins.A GHAIR explainer notes that researchers may increasingly weigh several factors at once, including biological activity, chemical stability, solubility, selectivity, metabolic properties, potential toxicity, and manufacturing feasibility, which could help them identify more balanced drug candidates earlier.
UCLA researchers report that their platform, MAESTRO, uses large language models to analyze scientific knowledge about drug mechanisms and cancer cell pathways. According to UCLA, MAESTRO evaluates drug combinations, ranks promising candidates, and generates biological explanations for testing, evaluating each candidate more than 100 times.
AstraZeneca says that for targeted protein degraders, its scientists combined high-throughput proteomics with machine learning, which helped them identify an improved molecule with a more favorable safety profile. For additional context on how OpenAI's models are evolving, consider recent discussions about models like GPT-4o.



