MedChemBench
Intro
MedChemBench is a medicinal-chemistry reasoning benchmark that OpenAI introduced on 20 November 2026. It tests whether models can connect a molecule’s structure with experimental evidence to explain results, compare compounds, and predict their properties. Official announcement: Introducing MedChemBench. The official @OpenAI account posted about the benchmark the same day, saying it shows how frontier models have improved at helping scientists with this work.
The official page says the benchmark has 99 problems drawn from published research, covering 117 papers and patents and 761 distinct molecular structures. 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 also notes that reconstructing structures and explaining evidence were relatively stronger, while selecting compounds and predicting numerical properties proved harder.
Advertisement
Explain
What problem it targets
A hard part of medicinal chemistry is that improving one property of a molecule can make another worse: a compound may act strongly against a target but be poorly absorbed or break down too quickly. Chemists work through those tradeoffs by designing, making, and testing related compounds, then deciding what to try next. For AI to help, a model needs to connect structural changes with experimental evidence and reason about which compounds are worth pursuing.
MedChemBench turns those judgments into scored problems. For most items, the model receives the context available at a chosen point in a research program: the biological target, the researchers’ goals, the compounds studied so far, and selected experimental findings.
What the official page says it grades
According to OpenAI, the benchmark evaluates:
- Molecular structures: correct atom connections and the specified stereochemistry;
- Scientific reasoning: explanations supported by the supplied evidence, such as why one compound was chosen;
- Comparing compounds and predicting numerical properties, which the official write-up says proved harder.
Limits
This is a benchmark, not a new generative model or a lab platform. The scores come from OpenAI’s official report under a stated reasoning setting and are not a discovery rate in a real laboratory. The problems are drawn from published papers and patents, so the test is reasoning from given historical evidence, not open-ended experimental work.
Usage
Who it is for
- Research and product teams that follow drug discovery, molecular design, or scientific-reasoning evals;
- People who need to compare frontier models on structure understanding, explanation, and compound decisions;
- Anyone who wants to cite the official numbers rather than a second-hand recap.
How to use it
- Start with the official write-up: Introducing MedChemBench;
- When you quote scores, keep the official setting: at xhigh, GPT‑6.2 Astra 84.5%, GPT‑6 Astra 83.3%, GPT‑6.1 Sol 82.0%, and GPT‑5.6 Sol 58.2%;
- Treat the benchmark as a test of reasoning from a given research context, not as evidence that a model can run a drug program on its own.
This entry is based on the official page and OpenAI’s same-day post.