AI is getting much better at a task that used to feel purely human: deciding which molecule to test next.
OpenAI announced MedChemBench on Thursday, a benchmark built from 99 medicinal chemistry problems pulled from published papers and patents. The dataset spans 761 distinct molecular structures. Models are given a biological target, a set of earlier compounds, and selected experimental findings as far back as a decision point. They then have to reconstruct the molecules, explain the evidence, choose which compounds to advance, and predict their properties.
The headline score is a percentage of available credit, not a perfect score. At the highest reasoning setting, GPT-6.2 Astra earned 84.5 percent. GPT-6 Astra followed at 83.3 percent. GPT-6.1 Sol reached 82.0 percent, and GPT-5.6 Sol landed at 58.2 percent.
The gap is real, but the profile of the scores is what matters. OpenAI's own paper separates the benchmark into four components. GPT-6 Astra earned about 95.4 percent of the credit for reconstructing molecular structures. It also scored strongly on scientific reasoning, explaining findings with the supplied evidence. But the same model earned only about half the credit for ranking and selecting compounds, and below half for predicting numerical properties.
This is the kind of result that does not make for a simple headline. A reader who only hears "AI scored 84.5 percent on a chemistry benchmark" will walk away with a false sense of what was achieved. The benchmark does not measure whether a model can discover a new drug. It measures whether a model can reason from historical evidence at a specific decision point in an already completed research program.
MedChemBench is a retrospective benchmark. Each problem stops at a decision point before the outcome is known. That design is intentional; it prevents the model from being handed the answer. It also means the task is not a full drug discovery workflow. The paper's authors were explicit about this. Every response in the study was a single attempt with no tools, no feedback, and no new experiments. Whether the model had seen the source papers before cannot be ruled out.
The paper also notes the benchmark is close to saturation in its current form. Open-ended design and synthesis grading still depends partly on reference answers, which can penalize plausible alternatives. That is why OpenAI describes the work as a step toward helping scientists, not a replacement for medicinal chemists.
One of the clearest examples in the paper is a problem built around PRMT5, a protein studied in cancer drug discovery. Models received molecular structures and earlier experimental results. They had to predict how three related compounds would behave across three properties: how strongly they block PRMT5 in cells, how readily they cross a cell barrier, and how quickly liver enzymes break them down. A compound can be strong against a target and still be poorly absorbed or metabolized too quickly. Choosing what to test next means weighing those tradeoffs at once, not optimizing a single measurement.
For PeDeDe, the decision to cover this story is not about the headline score. It is about the shape of the benchmark. The results show where AI is genuinely useful in science, and where the fallback is still human expertise. Structure reconstruction and evidence explanation are areas where the models are strong. Compound selection and numerical prediction remain comparatively weak. That is a useful, honest summary of a difficult technical result.
The coverage is timely. Antibiotic resistance and the cost of drug development have pushed medicinal chemistry toward AI-assisted tools for years. MedChemBench gives the field a shared standard, with open questions about generalization and retrospective bias that no headline number should smooth over.