How AI Solved a 1 Million Dollar Problem

OpenAI says its agents cracked part of the Navier-Stokes problem. Here is what it actually proved, and why the race to get there matters more than the math.

Sam Altman's company just said it broke a fluid dynamics problem that has resisted mathematicians for more than a century. The story got told online in the space of a single day, and most of what was said about it was wrong.

Here is the accurate version. OpenAI did not win the million dollar prize attached to this problem. It did not solve the exact question the prize is offered for. What it did do is still a genuine result, and the mess around how it was announced tells you something important about how AI companies are starting to compete on pure research, not just products.

Watch the full video breakdown on YouTube

Sponsored by AnyDB. Build flexible, connected operational systems around the way your business works, and connect them to Claude and ChatGPT through MCP. Try AnyDB

What Actually Happened

On September 6, OpenAI announced that a swarm of 10,000 of its AI agents had worked for 88 hours and produced a proof related to the Navier-Stokes equations, the set of formulas that describe how fluids move. Navier-Stokes is one of the seven Millennium Prize Problems, a list of the hardest unsolved questions in mathematics published by the Clay Mathematics Institute in 2000. Solve one, and the Institute pays out a million dollars. Only one of the seven has ever been solved.

The specific question the Millennium Prize asks is whether smooth solutions to the unforced Navier-Stokes equations, meaning no outside force is pushing the fluid, always stay smooth forever, or whether they can break down into a singularity in finite time. OpenAI's agents did not answer that question. They proved something narrower. For the forced version of Navier-Stokes, where something is actively stirring the fluid, they showed a singularity can occur. Separately, they resolved a related but easier question about the Euler equations, which describe fluid motion with no friction at all, for the unforced case.

OpenAI has said plainly that it is not claiming the million dollar prize, and the Clay Mathematics Institute has not verified or endorsed the work. Sébastien Bubeck, the OpenAI researcher who led the effort, said the push started after rumors circulated online that Anthropic might be close to a Millennium Prize result. His quote was direct: the team looked at its own model and decided to try.

That decision produced an awkward collision. Mathematician Tristan Buckmaster of New York University, working with Levent Alpöge, a researcher affiliated with Anthropic, had been making real progress on a closely related problem for weeks using an Anthropic model. When OpenAI finished its work, it reached out to Buckmaster and Alpöge to propose releasing both results together. The two teams had ended up proving different, overlapping pieces of the same broader puzzle, using different companies' AI systems, at almost the same time.

Why This Race Happened at All

Set aside the mathematics for a second and look at the business logic. Two AI labs raced each other to produce a hard math proof within weeks of each other, using models built for general reasoning rather than tools designed specifically for formal proof work. That is new. A year ago, a result like this would have taken a team of specialist mathematicians months or years, and no one would have called it a competitive event between two companies.

The reason this happened now is that proof checking has gotten cheap relative to proof generation. Once a mathematical claim is written in a formal language like Lean, a computer can verify every step mechanically, with no room for hand waving. That means an AI system can throw enormous amounts of compute at generating candidate proofs, and any wrong ones get caught automatically. OpenAI's own account credits this loop, thousands of agents generating attempts, an automated checker rejecting failures, for making an 88 hour timeline possible at all.

This changes the incentive structure for AI labs. A verified mathematical proof is a clean, publicly checkable demonstration of capability. It is harder to fake or spin than a benchmark score, and it is the kind of result that makes headlines outside the AI industry. Expect more of this, not less. If proof checking stays cheap, math becomes one of the clearest arenas for two labs to prove, in public, which one's model reasons better, and that turns hard problems into a marketing tool almost by accident.

The Lesson Here Is About Credit, Not Fluid Dynamics

The interesting failure in this story is not mathematical. It is about how credit gets assigned when two teams converge on overlapping results using different tools at almost the same moment. Academic mathematics has centuries of norms around priority, who submitted first, who peer reviewed what, who gets named on the paper. None of that was built for a situation where two commercial labs, each with something to gain from the announcement, are racing an unofficial clock.

Watch for this pattern outside math too. Any time an outcome depends heavily on which large model you point at a problem, and multiple companies have access to comparably strong models, you should expect near simultaneous, overlapping results to become normal rather than rare. The scarce resource used to be a brilliant specialist working for years. Increasingly, the scarce resource is compute and the willingness to point it at a well chosen target, and multiple companies can supply both of those at once.

Here's What I'd Do

If I ran a research team, I would not treat this as a reason to rush announcements before they are checked. OpenAI's own framing is the responsible part of this story: it published its methods and code, it did not claim the prize, and it acknowledged the Anthropic affiliated team's overlapping work instead of racing to bury it. That is the behavior worth copying, not the speed.

If I were reading about this as someone outside the AI industry, I would hold two things at once. The underlying capability is real and worth paying attention to, since AI systems generating and checking their own formal proofs at this scale did not exist a couple of years ago. At the same time, I would wait for independent mathematicians and the Clay Institute to weigh in before treating any specific claim as settled. Announcement day and verification day are not the same day, and with results this technical, the gap between them can be long.

Over to You

If two AI labs can now race each other to solve century old math problems in days instead of years, what fields do you think get shaken up next?