Everything About AI
NextFin News - OpenAI on Tuesday announced that an internal AI system had solved the Navier-Stokes existence and smoothness problem, one of mathematics' seven Millennium Prize Problems carrying a $1 million reward — but the claim landed just 12 hours after a New York University professor and an Anthropic researcher announced their own closely related result, setting off a dispute over scientific credit, training data, and who really got there first.
The Breakthrough and the Rival Claim
The OpenAI proof, produced by an internal model the company says is significantly more capable than GPT-6 Astra, shows that the dynamics of the Navier-Stokes equations for fluid motion can develop a singularity in finite time. The company published both a writeup of the proof and a formalization in Lean, the programming language used to mechanically verify mathematical arguments. The result resolves statement "C" (and "D") of the official Millennium Prize formulation set out by the Clay Mathematics Institute in 2000, a question that has remained open for roughly 90 years.
According to OpenAI, the effort began on September 1, after the company heard rumors that two Millennium Prize problems had been resolved. A group of on the order of 10,000 concurrent agents, subdivided and communicating within groups, arrived at the resolution on September 5 — about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours via GPT-6 Astra. Across all attempted problems, the agents sent 4.9 million messages and used about 300 billion output tokens; the Navier-Stokes effort alone accounted for 2.7 million messages and approximately 130 billion output tokens. At current Astra rates, that compute would cost about $22.5 million.
Twelve hours before OpenAI's announcement, Tristan Buckmaster, a mathematics professor at NYU, and Levent Alpöge, a researcher at Anthropic, announced findings on closely related problems, produced with help from Claude and OpenAI's Codex and built on earlier work by Diego Córdoba of the Institute for Mathematical Sciences in Madrid and Luis Martínez-Zoroa of CUNEF University. Their result stops short of a full Millennium Prize solution, but it moves in the same direction — and, critically, along the same uncommon route.
The coincidence is what makes the dispute more than academic gossip. Buckmaster says the pair had quietly chosen to attack the problem through a smooth force, options C and D in Charles Fefferman's official statement of the problem — the same path OpenAI's agents took. "Almost nobody else I know of was working on it," Buckmaster wrote. "It is not the direction one arrives at in a few days by giving a model the problem statement."
"It is not the direction one arrives at in a few days by giving a model the problem statement."
— Tristan Buckmaster, NYU mathematics professor
That last line is the crux. If the route is as uncommon as Buckmaster claims, independent convergence in a matter of days is statistically implausible — which is exactly why the timing of OpenAI's September 1 launch, after information about the academic work had already reached the company, sits at the center of the controversy. OpenAI's own post confirms the timeline: the effort started on September 1, inspired by the rumors, and the company reached out to Buckmaster and Alpöge on September 6, after Lean verification was complete, to propose a concurrent release that would recognize their priority.
The Data Question: Did OpenAI See the Work?
Buckmaster says that while he and Alpöge were finalizing their results, they learned that "information about our progress had been passed to OpenAI." When they pressed the company on when its research began and how much human input was involved, he says the answers became evasive. Eventually, he says, it was agreed the first prompt had been sent in the past few days — after information about their work reached OpenAI.
The deeper concern is whether OpenAI's internal model had access to the pair's private Codex sessions. Buckmaster and Alpöge relied primarily on OpenAI's Codex in their work, and OpenAI reserves the right to train models on Codex interactions, though users can opt out. "I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project," Buckmaster wrote. "I was told the model did not look up user data. I asked again, about training, and I did not get an answer."
OpenAI denies any improper access. In its blog post, the company said:
"We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem."
It added a caveat: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." Sebastien Bubeck, a member of technical staff at OpenAI, said the proofs differ significantly and "even the precise results proved are different in the Euler case (forced vs unforced)."
There is a second, sharper allegation, and OpenAI's side of it should be recorded. Buckmaster alleges that Bubeck asked him to remove Alpöge's credit as part of a proposed compromise, and that when Buckmaster pushed to make the dispute public, Bubeck replied, "Why would you ruin your career?" followed by, "If you don't want me to be nice, then I don't have to be nice." Bubeck has specifically refuted the claim that he asked for Alpöge's name to be removed. OpenAI's public statement does not address the conduct allegations directly.
Buckmaster responded that OpenAI is "openly admitting they used training data from a period after we found our result." He has said he does not know whether his data was actually used.
A Structural Shift in How Mathematics Gets Done
Whatever the outcome of the credit dispute, the event marks a structural shift in mathematical research rather than a one-off curiosity. Three ingredients came together: an unreleased frontier model with a step change in reasoning ability, a multi-agent system that could run thousands of parallel proof attempts, and Lean formal verification that removes the need for human referees to trust the argument.
The mechanism matters. For decades, the bottleneck in mathematics was not ideas but verification. A claimed proof of a Millennium Problem can take years for the community to check — Grigori Perelman's solution of the Poincaré conjecture consumed roughly four years of scrutiny. Lean changes that economics. If a proof formalizes, it is correct; the cost of verification collapses toward the cost of running the checker. That shifts the scarce resource from human referee time to compute. Labs that can afford 300 billion output tokens and 10,000 concurrent agents gain a durable advantage in the race for the deepest results.
This is not a cyclical fluctuation in research productivity that will mean-revert. It is a regime change in the production function of mathematics. The evidence: the verification technology (Lean) is permanent and improving; the compute scale is available only to a small set of well-capitalized labs; and the result has already been mechanically checked rather than resting on human consensus. A cyclical claim would require a short-term driver and a demonstrated mean-reversion pattern — neither exists here. The structural leg is the discovery engine; the cyclical leg, if any, is the near-term race to file claims on the remaining six Millennium Problems.
The distributional consequence is uncomfortable. Before this week, the frontier of mathematics was populated by researchers whose only constraint was their own time and insight. After this week, the frontier has a toll booth: the cost of running the kind of search that produces a Millennium-level proof is measured in tens of millions of dollars and clusters of GPUs that only a handful of entities control. The Córdoba–Martínez-Zoroa strategy that both teams built on came from academic researchers without frontier compute. They supplied the direction; the machine supplied the brute force. If that division of labor becomes the template, the intellectual rents will accrue to whoever owns the machine.
The Second-Order Consequence: The Prize Is Not the Point
The obvious first-order read is that OpenAI has solved a $1 million problem. That misses the stakes. OpenAI says it does not plan to take the prize money. The real prize is signaling value.
OpenAI is a private company racing toward a next funding round and, eventually, an IPO. A verified solution to one of mathematics' hardest problems, produced by a model "significantly more capable than GPT-6 Astra," is evidence that the next generation of models will be substantially more powerful — evidence the company can show investors without releasing the model itself. The $22.5 million compute tab is not a cost to be minimized; it is a down payment on a valuation narrative that says the frontier is still accelerating. In that framing, the $1 million prize is not the reward; it is the receipt.
The second-order transmission runs through the AI capital cycle. If frontier models can independently produce Millennium-level mathematics, the perceived runway for AI capability extends, which supports higher valuations for the compute suppliers and the labs that control the largest clusters. Nvidia, whose CEO Jensen Huang declared "AGI has arrived" on September 7 following the GPT-6 Astra launch, is the most direct beneficiary of that narrative — though its shares closed down 3.17% at $225.73 on September 8, a reminder that the market separates capability headlines from near-term revenue. Microsoft, OpenAI's largest backer, was essentially flat, up 0.19% at $493.95.
The muted equity reaction is itself the second-order story. Nvidia's 3.17% decline on the day of the announcement, on heavy volume of 118.4 million shares, suggests investors are no longer willing to bid up the entire AI complex on any capability headline. The market is asking the second question — not "can the model do it?" but "who captures the economics, and when does it hit revenue?" If the economics of discovery accrue to the model owners and the chip suppliers, the beneficiaries are narrow. If the cost of verification collapses for everyone, the value diffuses. Either way, the marginal dollar of AI enthusiasm now demands a revenue line, not a proof.
The Counter-Thesis: Extraordinary Claims, Ordinary Scrutiny
The strongest case against the triumphalist read is straightforward: the result is not yet accepted. The Clay Mathematics Institute has not awarded the prize, and formal verification in Lean, while powerful, does not guarantee the formalization matches the intended theorem. Independent mathematicians must still read the writeup and confirm the Lean statement encodes the actual Millennium Problem. Charles Fefferman of Princeton, who wrote the Clay Institute's official description of the problem, said he was "thrilled that the problem was solved" and called Córdoba and Martínez-Zoroa the heroes — but warmth from the problem's author is not the same as a prize committee's judgment.
There is a deeper counter-thesis, and it comes from one of the field's most respected voices. Terence Tao, widely regarded as one of the world's greatest living mathematicians, criticized the use of longstanding mathematical challenges as marketing proof points for AI prowess. Tao argued that by focusing on answers, AI discourages mathematicians from working on alternative approaches that might arrive at the same solution, and that AI companies rarely reveal the dead ends — which is precisely where the insights that open new fields of mathematics are found. He noted he had initially been hopeful that AI would be a wonderful tool in the hands of expert mathematicians, like a microscope for biologists or a telescope for astronomers, but that increasingly it was being used autonomously to produce answers without providing much insight. For Tao, the insight matters more than the answer. That is a direct challenge to OpenAI's framing, which presents the solved problem as evidence of capability rather than as a contribution to understanding.
There is also the credit problem. If a material portion of the approach was derived — even indirectly — from Buckmaster and Alpöge's unpublished work, the claim of independent discovery weakens, and with it the clean narrative that agentic AI alone cracked a 90-year-old problem. The conduct allegations are the sharpest edge, though they are contested: Buckmaster alleges that Bubeck asked him to remove Alpöge's credit, an allegation Bubeck has denied.
The falsifying signal is concrete: if the Clay Mathematics Institute declines to award the prize after formal review, or if independent mathematicians identify a gap between the Lean formalization and the official problem statement, the "AI has solved Navier-Stokes" claim is materially weakened. A second falsifier: if a forensic review established that the internal model had been trained on the pair's private Codex sessions, the independence claim would collapse. A third: if prominent analysts decline to endorse the formalization within weeks, the Perelman comparison — four years of scrutiny — will look less like caution and more like a warning.
What to Watch: Scenarios by Time Horizon
Short term (days to weeks): the mathematics community reads the writeup and the Lean code. Expect a flurry of blog posts and arXiv commentary. The signal to watch is whether prominent analysts — the equivalent of the Perelman-era verifiers — endorse the formalization.
Medium term (months): the Clay Mathematics Institute's review process. The institute has historically taken years to award Millennium Prizes; a faster or slower timeline will itself be a signal. Also watch whether OpenAI, Anthropic, and NYU reach any agreement on joint credit — a joint announcement would defuse the dispute; silence would let it harden. OpenAI, which closed a record $122 billion funding round at a post-money valuation of $852 billion in March, is widely expected to pursue a public listing, and the Navier-Stokes announcement lands inside that capital-raising window.
Long term (years): whether this becomes the template for frontier research. Base case: agentic systems plus formal verification become standard for the hardest problems, concentrating breakthrough production in a handful of labs. Upside case: the tools diffuse, and the cost of verification falls for university researchers, democratizing access to the frontier. Downside case: a high-profile retraction or a finding of data misuse chills collaboration between academic mathematicians and private labs, slowing the pipeline of human insight that these systems still depend on.
The base case carries a warning for the victors. The Córdoba–Martínez-Zoroa strategy — the intellectual foundation both teams built on — came from academic researchers without frontier compute. The machines finished the proof, but the direction came from humans. If private labs are seen as extracting and then outspending academic work, the trust that supplies them with ideas could erode. Tao's point lands here: a field that rewards answers over insight will get fewer insights, and eventually fewer answers.
Bottom line: the Navier-Stokes result is a genuine milestone for AI-aided mathematics, but the surrounding dispute is the more durable story — a structural shift in how discovery is produced, colliding with a governance gap in credit and data that the field has not yet solved.
Data as of market close September 8, 2026. OpenAI is a private company; equity figures refer to its publicly traded partners.
Explore more exclusive insights at nextfin.ai.
