An AI Chemist Improves a Stubborn Drug-Making Reaction in the Lab

OpenAI says one of its models, working as a near-autonomous lab agent, has improved a chemical reaction that medicinal chemists have wrestled with for decades. The result, published on June 17 from a three-month collaboration with the Polish startup Molecule.one, is being described as the first publicly documented case of a frontier AI model running a real wet-lab chemistry campaign and producing a change that held up under independent checking.

The target was the Chan-Lam coupling, a copper-catalyzed way of forming carbon-nitrogen bonds. Those bonds turn up throughout modern medicines, so a more dependable version of the reaction is genuinely useful rather than a curiosity. OpenAI connected its GPT-5.4 model to an automated chemistry platform that proposed conditions, ran experiments and read the results. Over the course of the project the setup carried out 10,080 reactions and landed on a TEMPO-based approach that improved a coupling chemists had long found temperamental.

Where the humans stayed in the loop

The word doing the heavy lifting here is "near." This was not a robot scientist left alone with the keys. Human chemists chose which of the model's proposals were worth trying, fixed flawed experimental plans, ran much of the bench work and independently confirmed the final result before anyone claimed it. The project also did not yield a new drug. It improved a step used to make drugs, which is a narrower and more honest claim.

Even with those qualifications, the shape of the work is notable. The model did more than answer questions. It proposed hypotheses, pushed them through thousands of physical experiments and interpreted what came back, the loop that scientific research actually runs on. That is a different kind of contribution from drafting text or summarizing papers, and it is the part worth paying attention to.

A pattern across the labs

The chemistry result fits a run of recent moves by OpenAI into hard science, from upgrades to its GPT-Rosalind drug-discovery model to a study, out the next day, in which its o3 model helped doctors solve rare disease cases that had gone unsolved for years. The thread tying them together is the automation of the slow, expensive search that sits underneath discovery, whether the haystack is the medical literature or a space of possible reaction conditions.

The open question is how far this generalizes. One improved reaction, validated once, does not remake chemistry. But a system that can run ten thousand experiments without losing patience, and now and then surface something a human missed, is a tool worth watching as it meets messier problems.