OpenAI’s AI Solved a 200-Year-Old Math Problem That Had Defeated Every Human Mathematician
On September 8, OpenAI announced that an unreleased AI model had solved the Navier-Stokes existence and smoothness problem — one of seven Millennium Prize Problems selected by the Clay Mathematics Institute in 2000 as some of the most important unsolved questions in mathematics. Each carries a $1 million prize. Before yesterday, only one of the seven had ever been solved.
It took OpenAI’s AI 88 hours. It had been unsolved for approximately 200 years.
What the Navier-Stokes Problem Actually Is
The Navier-Stokes equations describe how fluids move — water, air, blood, everything that flows. They underpin weather forecasting, aircraft design, ocean current modelling, and cardiovascular medicine. They are among the most practically important equations in all of science.
The problem, at its core, is this: do the Navier-Stokes equations always produce smooth, well-behaved solutions? Or can they break down — developing what mathematicians call “singularities” or “blowups” — in certain situations? The equations have been used successfully for centuries, but nobody knew whether they were mathematically complete or whether there were edge cases where they simply stopped working.
OpenAI’s proof showed that blowups do exist — that there are conditions under which the Navier-Stokes equations develop singularities and cease to describe the physics accurately. That’s the answer to the question. And it’s the answer that resolves the Millennium Prize.
How OpenAI’s AI Did It
The approach was unlike anything a human mathematician would attempt — and that’s precisely the point.
OpenAI first deployed approximately 1,000 AI agents working in parallel to solve a simplified version of the problem — the Euler regularity disproof — in roughly 50 hours. The agents sent nearly 5 million messages to each other during this process, coordinating their work across branches of the proof simultaneously.
Building on that result, OpenAI then deployed approximately 10,000 AI agents working concurrently on the full Navier-Stokes problem. After 88 hours, the agents produced a proof of a singularity in three-dimensional Navier-Stokes. An additional AI model then spent 17 more hours formalising the result — converting the proof into a machine-verifiable format that can be independently checked.
The agents had access to a cached version of the internet and the ability to run code. OpenAI stated explicitly that they did not access any unpublished work from other researchers, and that “no specific user data was accessed in order to solve this problem.”
The Credit Controversy
OpenAI’s announcement did not land without friction. Two mathematicians — Levent Alpöge of Harvard and Tristan Buckmaster of NYU — had been working on a related proof of a somewhat simpler version of the Navier-Stokes blowup problem. OpenAI acknowledged that it learned of their work on September 1 — through a rumour, not direct access — and intensified its own efforts as a result.
This has raised uncomfortable questions. Did OpenAI’s AI effectively race ahead of human mathematicians who were close to a solution, using knowledge that they existed to accelerate its own timeline? OpenAI offered Buckmaster and Alpöge concurrent release of results and full visibility into all prompts used, and said it did not see any of their work before their public release. Both researchers confirmed their work proceeded independently.
Nevertheless, the sequence of events — human mathematicians close to a solution, OpenAI learns a rumour, OpenAI deploys 10,000 agents, OpenAI publishes first — has prompted a genuine debate in the mathematics community about credit, priority, and what it means for human mathematical achievement when AI can outpace it on raw speed.
Why Mathematicians Are Both Excited and Concerned
The reaction from the mathematics community has been genuinely mixed — and that mixture is revealing.
The American Mathematical Society welcomed the announcement warmly, noting that the result builds on decades of human mathematical work by Navier, Stokes, Leray, Ladyzhenskaya, Córdoba, Martínez-Zoroa, and others. “This achievement, and the process that led to it, will bear fruit for a long time to come,” the society wrote. The proof doesn’t erase human mathematics — it extends it.
But Terence Tao’s concern is worth taking seriously. Mathematics isn’t just about producing correct answers. It’s about building human understanding — developing intuitions, frameworks, and ways of thinking that transfer to other problems. If AI can solve Millennium Problems in 88 hours but nobody fully understands why the proof works, the result is true but the understanding is absent. That matters for science in ways that go beyond any individual result.
The proof is currently being reviewed by external mathematicians. Formal verification by the second AI model is an important step, but independent human review of a proof this complex will take weeks or months.
What Comes Next
Six Millennium Problems remain unsolved. With OpenAI now deploying tens of thousands of AI agents on mathematical research, the question is not whether AI will attempt the remaining problems — it clearly will — but how quickly.
The Riemann Hypothesis, the P vs NP problem, and the Birch and Swinnerton-Dyer Conjecture are among the remaining six. Each is considered by mathematicians to be at least as hard as Navier-Stokes, and most are harder. But the same was said about Navier-Stokes two weeks ago.
OpenAI’s statement closed with a note of deliberate restraint: “This milestone represents substantial work by mathematicians and AI researchers. However, this is not a culmination, but rather a snapshot in time, of progress on AI development.” For more context on what OpenAI’s AI systems are now capable of, see our coverage of the GPT-6 Astra launch and the AGI era declaration and our earlier coverage of OpenAI’s frontier training pause.
Read more tech related articles here.
