GPT-5.4 and Molecule.one’s Maria system ran 10,080 reactions across two experimental cycles, identified TEMPO as a useful oxidant for a historically low-yield coupling reaction, and submitted a finding that four independent chemists supported as novel.
A drug molecule that cannot be synthesized is, for practical purposes, a molecule that does not exist. Medicinal chemists work inside that constraint every day. They design promising compounds on paper, but if the chemistry to build them produces low yields or messy byproducts, the molecules get abandoned and the program moves on. Synthesis is one of the quiet bottlenecks in drug discovery. That is the setting for a result OpenAI and Molecule.one published on June 17, 2026.
Working together, the teams connected GPT-5.4 to Maria, an agentic chemistry system tied to a high-throughput laboratory, and handed it an open-ended assignment: improve one of several important reaction classes. The system generated research proposals, designed and ran experiments, analyzed the data, and proposed follow-up work. Its most promising idea focused on Chan-Lam coupling, a reaction that forms carbon-nitrogen bonds appearing throughout the structures of medicines.
The gap the system targeted was specific. Coupling primary sulfonamides with boronic acids has historically given low yields, and that gap matters because sulfonamides appear in anticancer drugs, antimicrobials, and diuretics. GPT-5.4 narrowed the problem itself, identified primary sulfonamides as a difficult, high-value substrate class, then suggested that mild oxidants, TEMPO among them, could improve the reaction. The human chemists found the TEMPO suggestion unexpected, which is part of why it was worth running.
The proposal, labeled OAI-M1-03, was one of four selected from thousands the system ranked. Maria AI then translated the plan into detailed lab instructions, ran the experiments at high throughput, analyzed raw data, and returned structured results to GPT-5.4 for follow-up design. The full process ran three months, from March 4 to sharing results with independent experts on June 4.
Maria ran 10,080 reactions across the two cycles of OAI-M1-03 — more than a chemist running three reactions a day would complete in a decade. Mean yield rose from 16.6 percent to 25.2 percent. The share of reactions clearing 30 percent yield went from 15.6 percent to 37.5 percent. Measured yields improved for 88 percent of the boronic acids and 83 percent of the sulfonamides tested.
In the second round, the system found that TEMPO could be replaced by 4-hydroxy-TEMPO, a considerably cheaper analog, with little loss in performance. Human chemists then reproduced representative reactions by hand at bench scale and observed higher yields for 11 of 14 substrate pairs, with more than twofold improvement in eight of them. Four independent chemistry experts reviewed the preprint and supported the finding as novel. Tim Cernak at the University of Michigan described the result as a demonstration of mild conditions and a practical oxidant producing a broadly useful substrate scope.
The scope of what the paper establishes is worth stating directly. The work shows that a model can make a useful contribution to organic chemistry, proposing a specific surprising hypothesis, surfacing it for human review, designing experiments, interpreting data, and designing follow-ups. It does not show that AI can run a chemistry research program end to end. Human judgment stayed essential throughout. The workflow depended on specialized high-throughput infrastructure that few labs have.
Of the three other proposals the system generated and tested, two were experimentally confirmed and one was disproven, with analysis still ongoing. Independent replication and broader substrate characterization remain ahead. The next chapter belongs to the independent labs.
