OPENAI’S AI RAN 10,080 CHEMISTRY EXPERIMENTS AND ACTUALLY DISCOVERED SOMETHING THAT WENT INTO A REAL PAPER OpenAI and Polish chemistry startup Molecule.one published results from a three-month collaboration in which GPT-5.4 operated as a near-autonomous agent in a live pharmaceutical chemistry lab. This is not a simulation. The model proposed a hypothesis, helped design experiments, ran them through Molecule.one’s Maria Lab robotic platform, analyzed the results, and arrived at a finding that independent chemists described as novel. The specific chemistry involved Chan-Lam coupling reactions, a class widely used in drug discovery. Over 10,080 reactions, GPT-5.4 proposed using a TEMPO-based additive that no prior literature had applied to this reaction class. Under the optimized conditions, average estimated product yield jumped from 16.6 to 25.2 percent. The share of reactions clearing 30 percent yield went from 15.6 to 37.5 percent. These are the kinds of incremental gains that pharmaceutical researchers spend careers on. The fact that an AI model produced them autonomously in a 90-day partnership, working through a robotic lab system, is a meaningful shift. The finding is not a drug. It is a chemistry improvement that makes a common step in drug synthesis meaningfully more efficient. But the structure of how it happened matters as much as what it found. An AI agent identified a gap in existing knowledge, designed an experimental protocol to test it, executed at scale, and produced a result that passed peer review. That has not happened before. Keywords: OpenAI drug discovery, GPT-5.4 chemistry, AI pharmaceutical research, autonomous AI science