OpenAI's AI chemist improves a hard drug-making reaction

OpenAI describes a project with the drug-discovery company Molecule.one in which its GPT-5.4 model worked through a real chemistry problem and improved it. The target was Chan-Lam coupling, a copper-catalyzed reaction chemists use to build the carbon-nitrogen bonds that hold many drug molecules together. One version, coupling primary sulfonamides, has long given poor yields, which limits its use even though sulfonamides appear in more than 90 approved medicines.

GPT-5.4 was connected to Molecule.one's system, Maria, which controls a high-throughput automated lab. OpenAI says the model read the literature, generated and ranked research proposals, helped design experiments, and analyzed the results, while human chemists chose which ideas to test and checked the final answer. The lab ran 10,080 reactions and the system settled on a change using TEMPO that raised yields. When chemists then ran 14 representative substrate pairs by hand, 11 came out higher, and most more than doubled. The whole cycle took about two and a half months.

What is interesting is not that a model proposed an idea but that the loop closed: a hypothesis went into a physical lab, came back as data, and produced a result a human chemist validated. OpenAI's account of the work is careful to credit the chemists at every step.

Why it matters

If you work in drug discovery or lab automation, this is a concrete test of the "AI scientist" claim on a reaction people actually use. The honest read is mixed: it worked, but only with chemists steering every step and a robotic lab running thousands of attempts, so the bottleneck shifts to whoever has that hardware.