OpenAI tests an AI chemist in the lab

OpenAI published a June 17 study with Molecule.one in which GPT-5.4, connected to Maria, a chemistry AI integrated with a high-throughput lab, proposed an additive that improved a useful medicinal chemistry reaction. The target was Chan-Lam coupling applied to primary sulfonamides, a chemical motif found in many medicines. The verified result is concrete: under optimized conditions, yields improved for 88% of the boronic acids and 83% of the sulfonamides tested, and human chemists later confirmed improvement for 11 of 14 substrate pairs repeated at bench scale.

The point is not simply that the model produced an idea. The system generated research proposals, helped organize experiments, analyzed data and suggested follow-up work. Researchers still remained in the loop: they selected which proposals entered the lab, corrected some experimental plans, supervised laboratory operations and manually repeated representative experiments. OpenAI therefore describes the system as a near-autonomous AI chemist, not a human-free laboratory.

Why it matters

In drug discovery, synthesis is often a bottleneck. Teams can imagine many candidate molecules, but they can only evaluate them properly if they can make them with reliable enough yields. Chan-Lam coupling is used to form carbon-nitrogen bonds, which are common in active molecules, but this variant with primary sulfonamides has historically produced low yields. The identified additive, TEMPO and then a cheaper analog, does not by itself transform medicine discovery. It is more useful as a concrete example of how a model can help explore an experimental space that is too large for a conventional manual workflow.

The limits matter. The yield estimates first came from microliter-scale screening, the manual validation covered only 14 substrate pairs, and independent replication is still needed. OpenAI also emphasizes safeguards, because advanced chemistry capabilities can be misused. Still, this is worth watching because it moves the discussion about scientific AI away from broad claims and toward measurable results that specialists can review, reproduce and challenge.