OpenAI claims to have solved the 90-year-old Navier-Stokes Millennium Prize Problem in 88 hours. The company used a group of approximately 10,000 concurrent AI agents powered by a next-generation model described as significantly more capable than GPT-6 Astra.
The agents had access to tools for running code and reading from a cached version of the internet. They were organized into groups of varying sizes that could communicate internally. The process began on September 1 and the proposed solution was reached on Saturday, September 5.
The Clay Mathematics Institute, which offers $1 million for a solution to each Millennium Prize Problem, has not yet validated the proof. The mathematical community has not yet established whether the result is correct.
The achievement has led to accusations of research theft. Mathematicians Tristan Buckmaster from New York University and Levent Alpöge from Anthropic had been privately collaborating on the same problem. Buckmaster noted that OpenAI's approach appeared similar to theirs and was unlikely to be reached in a few days based solely on the problem statement.
OpenAI denied accessing private work to solve the problem, though it admitted that de-identified data from product usage could have helped improve its models. The dispute has brought renewed attention to OpenAI's July 29 program offering free access to frontier models for 10,000 researchers, with a goal of reaching 100,000 by 2027.