The OpenAI Navier-Stokes proof is a claimed solution, published on September 8, 2026, to one of the seven Millennium Prize Problems: OpenAI says an unreleased internal model, running as roughly 10,000 cooperating agents for about 88 hours, showed that a smooth fluid governed by the Navier-Stokes equations can blow up in finite time. On November 2, Sam Altman said that internal model is an upcoming GPT model, which is why the story is back in the news. The result is not yet accepted: the Clay Mathematics Institute says its review is "deliberately unhurried," and mathematicians at Cambridge and King's College London found that the written proof and its computer checked version do not fully match.
This guide covers what OpenAI actually claimed, how the run worked, the credit dispute with two mathematicians, what critics say, and what is known (and not known) about the next GPT model.
What OpenAI claimed about Navier-Stokes
The Navier-Stokes equations describe how fluids such as air and water move. They are used in aircraft design, weather forecasting and the study of blood flow. The open question, posed as a Millennium Prize Problem by the Clay Mathematics Institute in 2000, is whether a smooth three dimensional flow can develop a "singularity," where speeds grow without limit in finite time.
In its announcement, On the Navier-Stokes Millennium Prize Problem, OpenAI says its system produced both a written proof and a formal version in Lean (a programming language that lets a computer check each logical step). The claim, in OpenAI's words, is that "an initially smooth fluid at rest can develop a singularity in a finite time," with a smooth external force applied and finite energy throughout. OpenAI describes the shape of the solution as a vortex that spirals inward and stretches out "like spaghetti."
OpenAI says this establishes statement "C" (and also "D") in the official Clay formulation, which means it is a disproof of global smoothness, not a proof that solutions always stay smooth. That matches the direction many experts expected: before the announcement, Fields Medalist Terence Tao wrote on Mathstodon that the "emerging consensus" was that the answer to the regularity problem is negative.
OpenAI also says it will not claim the $1 million Millennium Prize.
How the internal model found the proof
According to OpenAI's write up, the work ran like this:
| Item | What OpenAI reported |
|---|---|
| Model | An internal model "significantly more capable than GPT-6 Astra," in training since August 28, 2026 |
| Setup | Groups of coordinating agents with code execution and a cached copy of the internet |
| Scale | On the order of 10,000 concurrent agents in the group that solved Navier-Stokes |
| Time | About 88 hours, from launch on September 1 to a result on Saturday, September 5 |
| Lean check | Another 17 hours of formalization and verification using GPT-6 Astra |
| Volume | 2.7 million messages and about 130 billion output tokens for Navier-Stokes |
| Whole effort | 4.9 million messages and about 300 billion output tokens across all problems tried |
| Side result | A disproof for the unforced Euler equations, from nearly 100 agents over about 50 hours |
OpenAI says it first pointed agents at every open Millennium Prize Problem, then shifted resources to Navier-Stokes after the Euler result came back. It used Codex to merge the most useful ideas between agent groups and upgraded the agents to a further trained version of the model partway through. It also says the run used "the same strict safeguards" as its other frontier model evaluations, including monitoring and isolation.
Those figures come only from OpenAI. Nobody outside the company has audited the run, and the model itself has not been released.
The credit dispute with Buckmaster and Alpöge
The announcement landed in the middle of a priority fight. On September 7, NYU mathematician Tristan Buckmaster posted a statement saying he and Levent Alpöge, an Anthropic employee, had proven finite time blowup with smooth forcing for the 3D incompressible Euler equations, plus two related fluid models. Buckmaster wrote that the work was a personal collaboration, that they used several language models including Anthropic's Claude and OpenAI's Codex, and that the core idea came from mathematicians Diego Córdoba and Luis Martínez-Zoroa.
Scientific American reported that Buckmaster alleged OpenAI had learned of their progress and adopted the same method. OpenAI's Sébastien Bubeck denied it, saying "We did not use their prompt or proofs to prompt our models." In a September 10 update to its post, OpenAI says an investigation confirmed Buckmaster's Codex prompts "could not have influenced the system in any way, including through training," and it credits the pair with priority on the forced Euler result.
Is the OpenAI Navier-Stokes proof verified?
Not in the sense that counts for the prize. Three separate checks are in play, and each has a different status.
| Check | Status as of November 3, 2026 |
|---|---|
| Lean formal proof | Published by OpenAI; a computer checks that it follows from its stated assumptions |
| Match between Lean and the written proof | Disputed: a Cambridge and King's College London preprint found places where they differ |
| Human review of the written proof | Ongoing; not peer reviewed in a journal |
| Clay Mathematics Institute | Said on September 11 the problem "has apparently been settled"; review is deliberately unhurried |
| Prize eligibility | Not yet possible under Clay's rules (journal publication, then two years, then general acceptance) |
The mismatch finding matters most. King's College London says the researchers found several places where the two versions do not say the same thing, including one where the written proof makes a stronger claim than the Lean version. New Scientist reported that the team believes the AI changed the argument while translating it so the Lean code would compile. Both sources are clear that the team has not shown the result is wrong. OpenAI told New Scientist it knows about the mismatch, that it does not make either proof invalid, and that it will fix errors in the written proof as they are found.
On the prize side, the Clay Mathematics Institute's rules say a solution must first appear in a qualifying outlet, then at least two years must pass, and then it must win general acceptance in the mathematics community. In its September 11 announcement, the institute said it shares the community's excitement but that the process "is deliberately unhurried."
Why the OpenAI Navier-Stokes controversy keeps growing
The fight is only partly about whether the math is right. The bigger argument is about what AI labs owe the field.
On October 6, OpenAI followed up with Sharing AI progress in mathematics, releasing a large batch of results from the same internal frontier model in its openai/math GitHub repository, with Lean formalizations for many of them. New Scientist put the count at 722 papers. OpenAI says the average result used compute roughly equal to three hours of ChatGPT Pro thinking, and on October 15 it added a $5 million contribution to a Simons Foundation program that helps mathematicians study AI results.
The Association for Human Mathematics answered with a statement saying "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power," and urged mathematicians to stop working with OpenAI. Tao's earlier warning set out the core worry: a solution produced mostly inside a company's AI harness, with the search process kept out of view, could settle a famous problem while adding little human understanding.
Pros and cons of AI solving open math problems
Pros:
- Results arrive in days rather than decades, and Lean versions can be checked by machine.
- OpenAI published the proof, the Lean files and some reasoning summaries rather than keeping them private.
- It gives a concrete, public signal of how fast frontier models are improving.
Cons:
- Lean only checks the formal version; the KCL and Cambridge work shows it may not match the written argument.
- Long machine written proofs are hard for people to read, so they may add little insight.
- Releasing hundreds of papers at once shifts the checking burden onto unpaid human reviewers.
- The search process ran on a private model that outsiders cannot test.
What we know about the next GPT model
Speaking at the ADIPEC conference in Abu Dhabi on November 2, Altman said OpenAI had solved one of the Millennium Prize Problems using 10,000 agents running "an upcoming GPT model" for about four days, working with "no hierarchy," according to The National. He also said "AI was mostly not as smart as our smartest scientists, and now it's mostly smarter than them," while insisting people must stay in control. Dealroom's report on the same remarks says he signaled the model could be released soon.
What OpenAI has not announced: a name, a release date, prices, plan access or a system card. Its October 6 post says only that it is "working to responsibly release the model that produced these results." Until then, the most capable OpenAI models you can use are in the GPT-6 family; our guide to GPT-6 Sol vs Astra vs Luna explains which one you get in ChatGPT and the OpenAI API.
Two things are worth watching when it ships. First, the system card: a model that ran 10,000 agents with little human steering raises the autonomy questions covered in how autonomous frontier AI models are. Second, the release pace: OpenAI's September post says it is "focusing on understanding this model" and may need "more deliberate choices about the pace of progress," the same debate behind Dario Amodei's call to pace the frontier.
What it means for you
- ChatGPT users: nothing changes today. The internal model is not in any plan yet.
- Developers: do not plan around a model with no published price or limits. Watch OpenAI's release notes.
- Students and researchers: treat AI math results as claims to check, not settled facts, until human review catches up.
- Anyone tracking AGI claims: this is strong evidence of fast progress in one domain, which feeds the debate in has AGI been achieved. It is not proof of general intelligence on its own.
Other labs' tools are part of this story too. Buckmaster's statement shows mathematicians already use Claude and Codex for research level proofs, and trackers such as Epoch AI follow how quickly frontier capability grows.
FAQ
Did OpenAI really solve the Navier-Stokes problem?
OpenAI claims it did, and the Clay Mathematics Institute said on September 11 that the problem "has apparently been settled." But no journal has published it, independent review of the written proof is ongoing, and a Cambridge and King's College London preprint found that the Lean version does not fully match the written one. Treat it as a strong claim that is not yet accepted.
What model did OpenAI use for the Navier-Stokes proof?
An unreleased internal model that OpenAI says is "significantly more capable than GPT-6 Astra" and had been in training since August 28, 2026. GPT-6 Astra was used only for the 17 hour Lean formalization step. On November 2, Altman described the internal model as an upcoming GPT model.
When is the next GPT model coming out?
OpenAI has not announced a date. Altman signaled at ADIPEC on November 2 that it could come soon, and OpenAI says it is working to release it responsibly, but there is no name, price or system card yet.
Will OpenAI get the $1 million Millennium Prize?
No. OpenAI says it does not intend to claim the prize. Clay's rules also require journal publication, a two year wait and general acceptance before any award is considered.
Why are mathematicians upset with OpenAI?
There are three complaints: a credit dispute with Buckmaster and Alpöge over the related Euler work (which OpenAI denies), the October release of hundreds of AI written papers at once, and the worry that machine produced proofs settle famous problems without adding human understanding.
