On September 8, OpenAI announced that an unreleased internal model—more capable than GPT-6 Astra—had solved the Navier-Stokes Millennium Prize Problem, one of seven unsolved math challenges with a $1 million bounty. The run used roughly 10,000 AI agents working in parallel for 88 hours, consuming an estimated 130 billion tokens at a cost analysts put above $40 million.
Two days later, OpenAI told The New York Times it had made "substantial progress" on a second Millennium Prize problem and was working out how to share the result. It did not name the problem. Rumor points to the Hodge Conjecture.
The controversy is not the math
Mathematicians are still reviewing OpenAI's Navier-Stokes proof. But the deeper issue is verifiability. OpenAI used a model the public cannot access, a process that cost more than most research labs' annual budgets, and formal verification in Lean—a proof assistant most working mathematicians do not read fluently.
NYU mathematician Tristan Buckmaster was working on the same problem and says OpenAI may have had access to unpublished work through ChatGPT logs. OpenAI denies it. Either way, the new capability is clear: if you can check an answer cheaply, you can now throw $40M and 10,000 agents at the search space. The math becomes reproducible only if you can afford the run.
What this means
Science has always required independent verification. When verification requires a model you do not control and a budget you do not have, peer review becomes pay-to-play. You can read the proof. You cannot re-run the search. You cannot audit the training data. You rent the conclusion.