A semi-autonomous AI chemist optimizes a medicinal chemistry reaction

OpenAI and Molecule.one used GPT-5.4 and the Maria AI to optimize a Chan-Lam coupling reaction, raising average chemical yields from 16.6% to 25.2%.

OpenAI details, with the company Molecule.one, an experiment where its GPT-5.4 model contributed to improving a reaction useful in pharmacy. Connected to Maria, [Molecule.one]'s agentic chemistry AI backed by a high-throughput laboratory, the model received an open-ended objective and carried out the steps in sequence: formulating hypotheses, designing and running experiments, analyzing data, proposing next steps. Chemists remained in the loop to guide, correct the plans, and manually validate.

GPT-5.4 independently targeted a notoriously difficult variant of the Chan-Lam coupling, which forms carbon-nitrogen bonds, by proposing an unexpected additive, TEMPO, a mild oxidant. Under optimized conditions, the study reports an average yield increased from 16.6 to 25.2%, and an improvement for over 80% of the tested substrates; a follow-up showed that a cheaper analog performed almost as well. In total, Maria performed 10,080 reactions, and chemists reproduced the results at the benchtop for eleven out of fourteen substrate pairs.

OpenAI describes the system as quasi-autonomous, with human judgment remaining crucial, and specifies that it neither establishes generalization to other reactions nor that an AI can independently lead a research program. Four external chemists judged the result unprecedented and noteworthy.