AI assembles molecules with 95.4% of points, but scores only about 50% on choosing drug candidates
OpenAI has introduced MedChemBench, an exam that tests AI on medicinal chemists' tasks. Given a molecule's structure and experimental results, the model must work out which of several similar compounds is better and what to synthesize next. The Astra model scored 95.4% of points for assembling molecules from drawings, but only about half for choosing compounds and predicting their properties. So AI can already be trusted to check structures, but picking a candidate for the lab is still best left to a chemist.
AI-processed from OpenAI Blog; edited by Hamidun News
Imagine a chemist has two very similar compounds. The first suppresses a disease-related protein more strongly but is poorly absorbed. The second is weaker but passes through the cell barrier better and breaks down more slowly in the liver.
Which one should be synthesized and tested next? These are exactly the decisions tested by OpenAI's new MedChemBench benchmark, a set of 99 tasks taken from real medicinal chemistry papers. The result is mixed: the Astra model earned 95.
4% of available points for assembling molecules from drawings and chemical fragments, but only about half for choosing compounds and predicting their numerical properties. For developers of AI tools in biotech, this means structure checking and data explanation can be automated now, but picking the "best candidate" cannot yet be handed to a model without a chemist's oversight.
How the test works
MedChemBench draws on 117 papers and patents and includes 761 distinct molecular structures. In most tasks, the model receives the same context the researchers had at a given step of the project: which target protein was chosen, what goals were set, which compounds had already been studied and what results had been obtained.
Four skills are assessed:
- Molecular structures. Connect atoms correctly and account for the spatial arrangement of groups (stereochemistry) in molecules that must be assembled from drawings and chemical fragments.
- Scientific reasoning. Explain, for example, why one compound was chosen over another, drawing on the data given in the task.
- Ranking and selection. Order or select compounds so that they meet the stated requirements.
- Property prediction. Give numbers and categories that match the measured and calculated values from the source. Answers close to the real measurement earn partial credit.
One task is based on a study of PRMT5 inhibitors. PRMT5 is a protein studied in the search for anticancer drugs. For three related compounds, the model must predict how strongly they block PRMT5 in cells, how easily they pass through the cell barrier and how quickly liver enzymes break them down in a laboratory test. Follow-up experiments, hidden from the model, showed a trade-off: compound 23 suppressed PRMT5 at a lower concentration, while compound 24 penetrated the barrier better and broke down more slowly.
What the numbers show
OpenAI ran eleven GPT models through the test at different reasoning-depth settings. In the xhigh mode common to all of them, GPT-6.2 Astra scored 84.5% of available points, GPT-6 Astra 83.3%, GPT-6.1 Sol 82.0%, GPT-6.1 Terra 75.1% and GPT-5.6 Sol 58.2%. In max mode, GPT-6 Astra reached 83.8%. Deeper reasoning generally improved results, but large gaps between models remained.
A telling example is a task to assemble a molecule from a central structure and two attached groups so that an 18-atom ring is preserved. GPT-6 Astra got it right in all ten answers. GPT-5.6 Sol and GPT-5.6 Luna failed every time: the first drew a 15-atom ring, and the second split the molecule into two unconnected fragments. In chemistry this is critical: if the model writes down the substance itself incorrectly, any further conclusions about its properties are meaningless.
The picture by skill is uneven. In structure assembly, Astra at xhigh scored 95.4% of points. It also explained experiments well: it noticed changes in structure and weighed their possible effect on binding to the protein, absorption and metabolism. But in choosing compounds and predicting numerical properties, it earned only about half the points. OpenAI cautions that these categories have different tasks and scoring rules, but the room for improvement is clear.
What this means in practice
A sensible division of labor right now looks like this: AI checks that a molecule is written correctly, helps interpret data and formulates hypotheses, while the chemist and the experiment decide which compound to make next.
OpenAI itself stresses the limitations. The test does not prove that models can discover new drugs or improve decisions in the lab. It cannot be ruled out that the models encountered the source publications during training. Each answer is a single attempt, without tools, feedback or new experiments. What was tested was reasoning on historical data, not actual drug discovery. Still, the industry now has a common yardstick that shows where models are already useful and where mistakes are still costly.
- Nov 20, 2026METR Updates Its Famous AI Chart: Task Length Doubles Every 4 Months, but Judgment Lags
- Nov 20, 2026OpenAI protects GPT-6.2 Astra's reasoning: theft success rate falls from 54% to 0.2%
- Nov 20, 2026Nvidia Is Close to Buying Inception for $500 Million to Get Diffusion LLMs
- Nov 20, 2026In China, Job Postings for AI-Exposed Professions Fell 10 Percentage Points Faster Than the Rest
Key Questions Answered
How the test works?
MedChemBench draws on 117 papers and patents and includes 761 distinct molecular structures. In most tasks, the model receives the same context the researchers had at a given step of the project: which target protein was chosen, what goals were set, which compounds had already been studied and what results had been obtained. Four skills are assessed: Molecular structures.
What the numbers show?
OpenAI ran eleven GPT models through the test at different reasoning-depth settings. In the xhigh mode common to all of them, GPT-6.2 Astra scored 84.5% of available points, GPT-6 Astra 83.3%, GPT-6.1 Sol 82.0%, GPT-6.1 Terra 75.1% and GPT-5.6 Sol 58.2%. In max mode, GPT-6 Astra reached 83.8%. Deeper reasoning generally improved results, but large gaps between models remained. A telling example is a task to assemble a molecule from a central structure and two attached groups so that an 18-atom ring is preserved.
What this means in practice?
A sensible division of labor right now looks like this: AI checks that a molecule is written correctly, helps interpret data and formulates hypotheses, while the chemist and the experiment decide which compound to make next. OpenAI itself stresses the limitations. The test does not prove that models can discover new drugs or improve decisions in the lab. It cannot be ruled out that the models encountered the source publications during training.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.