AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Source Summaries

On the Navier–Stokes Millennium Prize Problem (OpenAI, September 2026)

medium confidence · updated 2026-09-11

OpenAI post of September 8, 2026 reporting that an internal model 'significantly more capable than GPT-6 Astra', run as a system of coordinating agents, produced a proof that the three-dimensional Navier–Stokes equations can develop a finite-time singularity, establishing statements C and D of the Clay Mathematics Institute formulation. Records the method, agent counts, wall-clock time, token and message totals, the prior unforced-Euler result, and the concurrent Alpöge–Buckmaster work on forced Euler.

Contents

"On the Navier–Stokes Millennium Prize Problem" is a post published on openai.com on September 8, 2026 and bylined to OpenAI. It reports that an internal OpenAI system produced an analytical proof, with a Lean formalization, that a three-dimensional incompressible fluid starting at rest and subject to a smooth applied force can develop a singularity in finite time. OpenAI states this resolves the Navier–Stokes existence-and-smoothness problem by establishing statements "C" and "D" of the Clay Mathematics Institute's official formulation, and that it does not intend to claim the Millennium Prize.

The post is OpenAI's own account of the result and of the method that produced it. The mathematical content sits in two linked PDFs and a Lean repository, none of which is held as a raw source here.

The result claimed

The question at issue is whether smooth three-dimensional Navier–Stokes motion can break down despite viscosity, with fluid speeds growing without bound in finite time — an open problem the post dates to roughly ninety years, since Jean Leray's 1934 proof that solutions exist in a generalized sense, and one of the seven problems the Clay Mathematics Institute named in 2000.

The construction described is a vortex that spirals inward while elongating along its axis, with the central region shrinking and accelerating in a way that keeps total energy finite. OpenAI frames the technical difficulty as arranging for the equation's terms — acceleration, pressure gradients, momentum transfer, viscosity — to "both become _big_ yet _cancel_ in a precise way," so that the breakdown arises from the fluid's own motion rather than from an unbounded applied force, leaving the external force smooth as the velocity diverges.

Because statements C and D of the Clay formulation are the breakdown branches, the claimed resolution is a disproof of global smoothness rather than a proof of it.

Method

The post is the most detailed first-party account the record holds of a multi-agent system applied to an open research problem, and its operational figures are the part most likely to be cited.

  • The model. An internal model "significantly more capable than GPT‑6 Astra," in training since August 28, 2026, whose training was still ongoing and whose performance OpenAI says continues to improve. Agents were updated mid-effort when a further-trained version became available. See Astra.
  • Trigger. The effort began on September 1, 2026 after OpenAI "heard rumors that two Millennium Prize problems had been resolved," and was launched across all open Millennium Prize problems plus a few others.
  • Scale. Agents were subdivided into groups that could communicate internally; the Navier–Stokes group ran "on the order of 10,000 concurrent agents." Different groups received different variants of the problem statement, with versions A and B (proof branches) and C and D (disproof branches) assigned to separate groups.
  • Tools. Agents could read from a cached copy of the internet and run code. OpenAI states it maintained "the same strict safeguards that we apply to all our frontier model evaluations, including monitoring and isolation."
  • Cross-pollination. After an initial period, Codex was used to consolidate the most useful insights across agent groups into follow-up prompts drawing on the agents' own intermediate results; the group that reached the solution was guided this way.
  • Timing. The agents reached the resolution on Saturday, September 5, 2026, about 88 hours after launch. Lean formalization and verification took a further 17 hours, run via GPT‑6 Astra.
  • Consumption. Across all attempted problems, 4.9 million agent messages and about 300 billion output tokens. For Navier–Stokes alone, 2.7 million messages and approximately 130 billion output tokens.

A precursor result is reported alongside: on an "easier" problem set, nearly 100 agents working for about 50 hours produced a disproof of regularity for the unforced Euler equations — the zero-viscosity limit of Navier–Stokes. Seeing that result is what led OpenAI to shift agents off the other Millennium problems and prompt them with the Euler resolution.

The priority dispute

OpenAI's account of the concurrent work is that the rumor which triggered the effort related to Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a mathematics professor at NYU. On September 6, 2026, after completing its own project and Lean verification and believing the pair also had a Navier–Stokes solution, OpenAI contacted them to offer a concurrent release and recognition of priority in a joint announcement, and says it then learned their result was a resolution of the forced Euler problem. OpenAI states it offered them visibility into all prompts used and later into the proof, and writes: "We recognize the priority of their work on forced Euler and congratulate them on their remarkable mathematical achievement."

On data access the post states that neither the researchers nor the agents saw any of the pair's work before public release and that no specific user data was accessed, while adding that OpenAI "cannot rule out that de-identified data derived from their usage of our products helped improve our models," and that the proofs and even the precise results differ — forced versus unforced in the Euler case.

Two further details entered the record on September 11, 2026. Sam Altman said the attempt followed "rumors on the internet last week that Anthropic's models had solved a millennium problem," and the Alpöge–Buckmaster effort was described as having run for months using both Claude and Codex. Buckmaster published his communications with Bubeck and raised whether OpenAI used his data to solve the problem; Bubeck denied what he called Buckmaster's "false and inflammatory allegations," and OpenAI said that "no specific user data was accessed in order to solve this problem" (Source: transformernews.ai).

The characterization of the Alpöge–Buckmaster result is not stable across sources. This post describes it as a resolution of the forced Euler problem; Sébastien Bubeck is quoted in Wired's contemporaneous account referring to "their work on unforced Euler," which is the variant OpenAI's own agents claim (Source: wired.com). Wired reported that some academics disputed the credit framing, and Axios reported the same day that the result was "overshadowed by credit controversy" (Source: axios.com).

Buckmaster made a specific allegation about the exchange. He and Alpöge posted documents claiming advances in an area relevant to Navier–Stokes on September 7, 2026; on September 8 it became public that Buckmaster alleged OpenAI learned of their work, directed significant resources at the problem, and then proposed that he publish a paper announcing the Navier–Stokes result without Alpöge's name. Buckmaster said that when he threatened to make the exchanges public, Sébastien Bubeck responded, "Why would you ruin your career?" Bubeck denied suggesting Alpöge's removal, and OpenAI executives said in a press briefing that neither researchers nor agents saw the pair's work before publication (Source: wired.com; axios.com).

OpenAI head of research Mark Chen put the compute cost of the effort "in the millions of dollars," and the Millennium Prize OpenAI declines to claim carries a $1 million award (Source: wired.com).

The Clay Mathematics Institute had not accepted the proof. Institute president Martin Bridson said evaluation would be deliberately unhurried and rigorous, and the institute still listed Navier–Stokes among its unsolved Millennium Prize problems as of September 9, 2026. The same account records that Bubeck estimated the computing cost at several million dollars, consistent with Chen's figure, and that a roughly $15 million figure discussed at the press conference was a customer price for the same job rather than OpenAI's own bill. It also traces the cascade technique underlying the construction to Luis Martínez-Zoroa's 2021 dissertation and his later work with Diego Córdoba, neither of which used computers (Source: implicator.ai).

Framing

OpenAI presents the release as reporting on the pace of its own progress rather than as a mathematical claim to be prized: "Our goal in releasing this result is to report on the substantial progress of our AI models. We do not intend to claim the Millennium Prize for this result." It cross-links its September 6, 2026 disclosure "Research acceleration: The view inside OpenAI" for the proposition that "we are now in the next period of AI progress," and closes on pacing: OpenAI says it is "focusing on understanding this model" and that building steerable and accountable systems "may require more deliberate choices about the pace of progress." The same pacing language appears in chief scientist Jakub Pachocki's "An Alien Mind" of September 6.

Provenance

Captured in full from openai.com, the publishing organization's own domain, with body text, footnote and outbound links preserved and site chrome removed. Three artifacts named in the post are not held here: the Navier–Stokes proof PDF and the Euler proof PDF, both on cdn.openai.com, and the Lean formalization repository at github.com/openai/NavierStokesAndEuler. As of September 9, 2026 no independent verification of the claimed proof by mathematicians outside OpenAI is recorded here, and the internal model used is not named or otherwise documented.

Relationships

Categories: Source Summaries | global | frontier-models | technical | company | developer | ai-for-science | safety
What links here (8)