OpenAI announced on Tuesday, 8 September 2026, that it has found a solution to a decades-old advanced maths problem using a new artificial intelligence (AI) model and thousands of AI bots. The ChatGPT-maker stated that a group of approximately 10,000 AI agents, which undertake tasks somewhat autonomously, solved a notoriously difficult maths problem in just 88 hours.
The problem is part of the Navier-Stokes equations, which concern how fluids move, and for 90 years, important aspects of these problems have lacked a proof. OpenAI referred to its solution as a "milestone" and evidence of rapid improvements in AI tools.
However, OpenAI's solution has yet to be independently verified or publicly accepted by The Clay Mathematics Institute, a US-based maths organisation that offers a Millennium Prize for solving certain mathematical conundrums. The solution reportedly resolved two out of the four statements required for the Millennium Prize proof, which is worth $1m to a winner.
The company stated it began training a new model at the end of August, which quickly demonstrated mathematical aptitude. OpenAI decided to use this internal model on advanced maths problems after hearing rumours on 1 September that two Millennium Prize problems had been resolved. By 5 September, 88 hours after setting 10,000 AI bots to the task, OpenAI claims to have found a solution to the Navier–Stokes existence and smoothness problem.
The claim has already generated controversy. Tristan Buckmaster, a mathematics professor at New York University, stated on Tuesday that he and Levent Alpöge, a mathematician at OpenAI rival Anthropic, had also been working on solutions to the problem using OpenAI's tool Codex. Buckmaster alleged that information about their progress was passed to OpenAI on 3 September, and that OpenAI did not begin working on the Navier–Stokes equations until after this information reached them. OpenAI congratulated the "concurrent work" of Buckmaster and Alpöge, stating it had not seen their work until its public release and that no user data was accessed, though it could not rule out that de-identified data from their product usage helped improve its models.