The week of September 8, 2026, has a subtle, unnerving quality. A bank of servers running 10,000 AI agents silently concluded that human mathematicians had been pursuing for the better part of a century somewhere in the background of everyday life, including morning coffee, traffic, and the usual noise.
The Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems chosen in 2000 by the Clay Mathematics Institute as the most challenging unsolved mathematical problems, was declared solved by OpenAI. The reward? A million dollars. The response from the mathematical community? complicated.
Everything from the drag on an airplane wing to the flow of blood through a coronary artery is described by the Navier-Stokes equations. These equations, which have their roots in the work of Claude-Louis Navier and George Gabriel Stokes in the 19th century, have proven to be both consistently helpful and obstinately enigmatic.
Whether fluid motion described by these equations could theoretically develop a “singularity”—a point where velocity grows to infinity in finite time, effectively breaking the model—was the main question. The smooth version of that question remained unanswered for almost a century after Jean Leray demonstrated in 1934 that solutions exist in a generalized sense.
According to reports, OpenAI’s system, which is said to be much more capable than GPT-6 Astra, proved that a singularity can indeed form. A vortex, a spinning swirl of fluid, spirals inward and elongates like pulled taffy, a narrow region speeding up while somehow maintaining its total energy as finite, according to the mechanism it discovered. The 166-page proof of this odd and counterintuitive result has been formalized in Lean, a programming language that verifies mathematical logic step by step. It is currently being reviewed by the Clay Mathematics Institute. Notably, that review is still ongoing.
The awkward economics of the entire endeavor are difficult to ignore. Noam Brown, a researcher at OpenAI, freely admitted that the effort cost millions of dollars—more than the prize it was pursuing. By pointing out that earlier AI benchmarks, which used to cost half a million dollars to run, can now be attempted for twenty dollars a month, he presented this as a positive indication of declining costs over time. Even though it sounds a little odd coming from a company that recently spent more money solving a prize problem than the prize is worth, there is logic to that argument.
The response from the math community has varied from cautious curiosity to complete skepticism. Richard Borcherds, a Fields Medal winner at UC Berkeley who studies abstract algebra and spends his time considering some of the most challenging problems in the field, stated that he had anticipated a result like this in roughly ten years, not right now.
That’s a quiet alarm as well as a compliment. Additionally, he stated that AI is already more capable than first-year graduate students and that this gap will only widen, with a bluntness that somehow feels refreshing. Berkeley graduate students seem to be aware of this. A few have begun to drop out.
Tristan Buckmaster, a mathematician at NYU, expressed a completely different kind of worry prior to publication, claiming that OpenAI might have expanded on the work he and his associates had completed, possibly using data processed through OpenAI’s Codex system. Additionally, he claimed that Levent Alpöge, a mathematician, was not given credit due to his affiliation with Anthropic, OpenAI’s primary competitor.
The data claim was rejected by OpenAI. It is likely impossible to determine whether the denial is true because, as researcher Camillo De Lellis of the Institute for Advanced Study noted, language models are opaque even to those who operate them. When attempting to establish scientific priority, that is an uncomfortable place to be.
The actual contribution of those 10,000 agents is also unknown. Noam Brown asserts that nobody truly knows. There was never a systematic comparison of the computation’s scale to what a single agent might have accomplished with more time. It’s possible that the solution would have been found at a much lower cost. The swarm might also have been crucial. To find out, no one experimented.
There’s a sense that mathematics is truly at a turning point at the moment, and those who are most passionate about it—graduate students, postdocs, and professors with years of experience—are attempting to recalibrate in real time. Richard Borcherds has begun to claim that AI provides him with superior answers to complex mathematical problems compared to his human colleagues.
This doesn’t seem to bother him. Others obviously are. As she completed her Ph.D. at Berkeley, Robin Huang described the entire endeavor as “extremely cynical,” pointing out how AI companies have fostered a climate of fear and rivalry where none previously existed.
The Navier-Stokes announcement has demonstrated that the rate of advancement in AI is no longer keeping up with human intuition about what is possible, regardless of the Clay Institute’s final decision. It’s still very much up in the air whether that’s something to celebrate.
FAQs
1. What did OpenAI actually prove about the Navier-Stokes equations?
Smooth fluid motion can develop a velocity singularity — breaking down the continuum model — in finite time.
2. How much did it cost OpenAI to solve this million-dollar problem?
The solution cost millions of dollars, exceeding the prize it was competing for.
3. How long did OpenAI’s AI system take to crack the problem?
10,000 AI agents working together solved in just 88 hours.
4. Why are mathematicians skeptical despite the formal Lean proof?
The Clay Mathematics Institute hasn’t confirmed the result, and key credit disputes remain unresolved.
5. What does this mean for math graduate students and researchers?
Many are reconsidering their futures as AI now outperforms beginning graduate students in mathematics.
