OpenAI's Navier-Stokes Proof Comes With a Credit Fight Attached

OpenAI says an internal, unreleased model produced a formally verified proof for one of mathematics' seven Millennium Prize Problems, and the openai navier-stokes proof announcement immediately got tangled up in a credit dispute with two mathematicians who say the company raced to beat their unpublished work to the finish line.

On September 8, OpenAI published its own account of the result: a coordinated system of roughly 10,000 concurrent agents, run on an internal model more capable than GPT-6 Astra, produced an analytical proof and a Lean-verified formalization showing that a smooth, finite-energy fluid in three dimensions can develop a singularity in finite time. That resolves the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems the Clay Mathematics Institute named in 2000, each carrying a $1 million prize. Only one of the seven had been solved before OpenAI's claim. OpenAI is explicit that it is not seeking the prize itself, and it has not yet been formally verified by the mathematics community, which typically takes months for a result this significant.

How the multi-agent system got there

The mechanics, laid out in OpenAI's own writeup, are worth more attention than they got in the day's headlines. The effort started on September 1, when OpenAI heard rumors that two Millennium Prize problems had been solved elsewhere and decided to test its newest internal model against all seven, plus a handful of other hard problems. Agents were split into groups, each prompted with a different variant of a problem statement (versions "A" and "B" of Navier-Stokes aimed at a proof, versions "C" and "D" aimed at a disproof), and allowed to communicate within their own group but not across groups.

A smaller group, under 100 agents, resolved an easier related question, the unforced Euler regularity problem, in about 50 hours. That result redirected the whole effort: OpenAI shifted its agents onto Navier-Stokes and fed them the Euler solution as a starting point. Periodically, OpenAI used Codex to cross-pollinate the freshest insights across agent groups rather than letting each one work in isolation. The winning group reached its result on September 5, about 88 hours after the first agents launched, with Lean formalization and verification taking another 17 hours. Across the full effort, the agents sent 4.9 million messages and burned through roughly 300 billion output tokens; the Navier-Stokes result alone accounted for 2.7 million messages and 130 billion tokens.

That architecture, many parallel agent groups working divergent framings of the same problem, a periodic consolidation pass to merge the best partial insights, and a willingness to redirect the whole swarm the moment one branch found traction, is the more durable takeaway of this story for anyone building agentic systems. It is a research pattern, not a math trick, and it will show up again on problems that have nothing to do with fluid dynamics.

Where the credit dispute comes from

OpenAI's own post acknowledges the awkward part directly. The September 1 rumor that triggered the whole effort turned out to trace back to Tristan Buckmaster, a mathematician at NYU, and Levent Alpöge, who works at Anthropic. The pair had spent most of a year making slow progress on a related question before accelerating sharply in mid-August using Claude and OpenAI's own Codex, running on GPT-5.6 Sol, to push an existing line of attack from mathematicians Diego Cordoba and Luis Martinez-Zoroa through to completion. Buckmaster has called the acceleration "a Deep Blue-Kasparov moment for mathematics," in a statement he released before OpenAI's announcement, reported in detail by Fortune and TechCrunch.

By Buckmaster's own account, he reached out to OpenAI first, on September 3, after hearing that a rumor about a solved Millennium Prize problem might trace back to his and Alpöge's unpublished work. He asked to speak the following week; OpenAI's team pushed to meet sooner, and two calls happened on September 6 with researcher Sébastien Bubeck on the line. On those calls, he says it emerged that a full OpenAI team had been working the problem, that the first prompt had gone out only in the days after word of his and Alpöge's progress reached the company, and that OpenAI's model had been pointed at the same forced-Navier-Stokes route he and Alpöge had quietly chosen. He also says he asked directly whether the model had been trained on or had access to their Codex sessions and got no clear answer on the training question. Two proposals came out of the calls, by his telling: post his Euler result and OpenAI's Navier-Stokes result on the same or adjacent days, with OpenAI publicly calling the pair the "closest humans to the problem," or have Buckmaster alone write up the Navier-Stokes result while acknowledging OpenAI's model had solved it. Separately, he says Bubeck twice pushed to have Alpöge removed from authorship over his Anthropic affiliation. Buckmaster declined both proposals and said he'd make the exchange public; he says Bubeck replied, "why would you ruin your career?" OpenAI has not disputed that specific exchange in public; its own post instead says the researchers and agents "did not see any of their work through any means until they released it publicly," while acknowledging, "while unlikely, we cannot rule out that de-identified data derived from their usage helped improve our models." No independent account of the private calls has surfaced, so that part of the story rests on Buckmaster's telling.

OpenAI's own post adds a specific and consequential admission: "while unlikely, we cannot rule out that de-identified data derived from their usage helped improve our models," even while it maintains no researcher or agent saw Buckmaster and Alpöge's actual work before it was published. That hedge, buried in a section titled "Concurrent work," is the detail worth remembering longer than the personal back-and-forth.

What it means for anyone running real work through an ai agent

Set the personalities aside and there is an operational lesson here for anyone doing genuinely novel or competitively sensitive work inside a coding or research agent. OpenAI's own language, "cannot rule out," is not a denial. It is an acknowledgment that a company cannot fully trace what its model may have absorbed from de-identified product usage, even when it is confident no person or agent directly viewed the underlying work. If you are running proprietary research, an unreleased product design, or anything you would not want a well-resourced competitor racing to replicate through a coding or research agent, that is the sentence to sit with, not the CEO drama around it.

There is a second, quieter thread here from mathematician Terence Tao, who argued before this announcement that AI systems optimized to produce answers to open problems, rather than to work alongside a human toward one, risk destroying the very insight that makes a field advance, comparing indiscriminate problem-solving by AI to strip-mining an archaeological site for artifacts while destroying the context that made them meaningful. That critique generalizes past mathematics: an agent that gets you the answer fast is not the same thing as one that helps you understand why the answer is true, and the two are easy to conflate when the output looks identical.

Sources: OpenAI, "On the Navier-Stokes Millennium Prize Problem"; Fortune; TechCrunch; Tristan Buckmaster's statement.