OpenAI Says Its AI Solved a 90-Year-Old Math Problem. The Mathematicians Who Actually Solved It Aren't Happy.
OpenAI announced today that an internal AI system produced a solution to the Navier-Stokes existence and smoothness problem — one of the seven Clay Millennium Prize problems, open since the 1930s. The announcement came with a full proof write-up and a Lean-verified formalization. It should have been an unambiguous win for AI-assisted math.
Instead, it kicked off one of the messier disputes the AI research world has seen this year.
What Navier-Stokes Actually Asks
The Navier-Stokes equations describe how fluids move — water in a pipe, air over a wing, weather in the atmosphere. Mathematicians have used them for two centuries, but nobody could prove whether the equations always behave: does a smooth, well-behaved fluid flow in 3D ever spiral into a singularity in finite time, where the math itself breaks down?
Proving it either way — that blow-ups never happen, or finding one that does — has been an open Millennium Prize problem since 2000, with $1 million attached and roughly 90 years of failed attempts behind it.
OpenAI's system claims to have found a genuine blow-up solution: an initially smooth flow that develops a singularity in finite time. If it holds up, that's a real answer to a 90-year-old question.
The Part OpenAI Didn't Announce
Here's where it gets complicated. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge say they'd already been closing in on essentially the same result — three separate finite-time blow-up proofs, including one for a related equation, using a distinctive mathematical approach they'd developed over months.
Buckmaster's public statement alleges that details of their unpublished work reached OpenAI, which then threw significant compute at catching up along their exact line of attack. OpenAI's own math lead, Sebastien Bubeck, has denied any wrongdoing. Sam Altman weighed in directly too, saying the Anthropic-affiliated side only had a narrower result (the Euler equations, not the full Navier-Stokes case), that coordination between the two teams broke down, and that his team was threatened with plagiarism accusations.
Nobody disputes the underlying math is hard and the results are real. What's contested is who got there first, how, and whether "the AI solved it" quietly depends on knowing which direction to point a very expensive model.
Why the Cost Question Matters More Than It Sounds
AI researcher Noam Brown put a number on it: this result reportedly cost millions of dollars in compute. He also made the obvious follow-up point — when OpenAI's o3 model first showed off strong reasoning results, it cost roughly $50 per task and looked absurd. That cost fell off a cliff within a year.
That's the pattern to actually watch here, more than the dispute itself. A "millions of dollars to solve one Millennium Prize problem" result is a stunt today. If costs keep falling the way they have for every other frontier capability, it stops being a stunt and starts being a research method — deploy a swarm of agents at an open problem, in a field where humans have made no progress in decades, and see what comes back.
The Uncomfortable Incentive Problem
Whatever actually happened between the two teams, the dispute exposes something real: once AI labs can plausibly race toward the same unpublished proof a human researcher is close to finishing, "who talks to whom, and when" becomes a real competitive and ethical fault line — not just a courtesy issue. Career-defining math results used to be safe from this kind of pressure. That may not be true anymore.
What This Means If You Use OpenClaw
The actual capability underneath this story — AI systems working a genuinely hard, open-ended problem largely on their own, over an extended stretch, checking their own work — is exactly the shape of what agents are for. OpenClaw is built around that same idea, just aimed at practical work instead of Millennium Prize problems: an agent that can take a task, run with it autonomously, use the tools it needs, and keep going without you babysitting every step.
You don't need a research lab's compute budget to see what agentic AI can actually do. Pick a tutorial on ClawWorld, point your agent at a real task, and watch it work the problem the way these systems are increasingly built to: independently, with memory of what it's already tried.