The counterintuitive additive

Chan-Lam coupling stitches a carbon-nitrogen bond onto sulfonamides, a chemical fragment that shows up in more than ninety approved drugs. It is also cranky: yields on primary sulfonamides run low, and a lot of the starting material simply falls apart. Over a roughly three-month collaboration, OpenAI's GPT-5.4 and Molecule.one's chemistry agent picked this reaction as their target, then drove an automated wet lab through 10,080 microliter-scale reactions looking for a way to make it behave.

Use of TEMPO in Chan-Lam coupling is underexplored and for the first time broadly tested here. — the preprint authors

The answer was a stable radical called TEMPO. Chemists mostly reach for TEMPO to mop up stray radicals, so proposing it as a *helpful* additive here is close to backwards — but it worked by blocking a side reaction that was destroying the boronic acid before it could couple. Adding it raised the mean estimated yield from 16.6% to 25.2% — a jump of about half in relative terms, though only some nine percentage points in absolute yield — and better than doubled the share of reactions clearing a usable 30%. At bench scale, 11 of 14 substrate pairs beat the old recipe.

What makes this a marker rather than a stunt is which part the model did. Earlier autonomous labs executed a procedure a human specified or optimized inside a fixed box; here the frontier model chose an open-ended problem from the literature and steered it to a non-obvious, transferable chemical insight. The humans were still firmly in the loop — writing the steering prompts, approving the runs, doing the bench validation, writing the paper — and the vivid before-and-after carries an asterisk, since microliter yields read low next to full-scale runs and OpenAI released the protocols but withheld the raw high-throughput data. The point is not that a robot did chemistry. It is that a general-purpose model, given no recipe, chose the problem and surfaced a fix the specialists had left on the table.