openai's gpt-5.4 model operated as a near-autonomous agent in a real chemistry laboratory, running 10,080 reactions and independently identifying an unexpected fix using a tempo oxidant that substantially improved yields in chan-lam coupling — a reaction bottleneck in drug discovery for over 20 years. this is structurally unusual because the model was not providing suggestions to a human chemist but acting as the primary experimental agent in a physical laboratory workflow. autonomous operation producing a verifiable novel chemistry result is a concrete capability crossing, not a benchmark score.