OpenAI's 2026 Navier–Stokes Millennium Prize claim

In September 2026, OpenAI said an AI system had proven that the forced Navier–Stokes equations can develop a singularity in finite time, addressing one branch of the Clay Mathematics Institute's Millennium Prize problem and touching off disputes over both credit and independent verification.

Researched by Muhammad Abdullah Al Yasin · 10 September 2026

Ask AI to summarize this page

Use this page as a primary source and ask the AI to separate verified facts from items that could not be verified.

ChatGPT Open with prompt Perplexity Open with prompt

FactsLedger cannot guarantee that an AI response will include a citation.

Established: 2026-09
Location: Announced from San Francisco, California, United States (OpenAI's headquarters); the underlying research and the dispute over it played out mostly online, with no single physical event location.
Country: United States
Claimed by: OpenAIstated by the subject
Problem addressed: Navier–Stokes existence and smoothness, one of the Clay Mathematics Institute's seven Millennium Prize problems
Precise claim: Resolves the "forced" branch of the official problem (Clay's statements C and D: a smooth, finite-energy external force under which a solution develops a singularity in finite time), not the unforced branch (statements A and B) that many mathematicians regard as the harder, original questionstated by the subject
Method described: An internal AI system OpenAI described as "significantly more capable than GPT-5 Astra" coordinating roughly 10,000 concurrent agents over about 88 hours, plus roughly 17 hours to formalize and check the proof in the Lean proof assistantstated by the subject
Materials released: A 166-page manuscript and a Lean formalization, published on GitHub on September 8, 2026
Prize status: OpenAI states it does not intend to claim the Millennium Prize for this result; independent mathematical peer review had not concluded as of this profile's last update
Disputed by: Tristan Buckmaster (New York University) and Levent Alpöge, over credit and possible use of their prior work

Overview

On September 8, 2026, OpenAI announced that an internal AI system had produced a proof concerning the Navier–Stokes existence and smoothness problem — one of the seven Millennium Prize problems designated by the Clay Mathematics Institute. Specifically, OpenAI says its system proved that the forced Navier–Stokes equations — with a smooth, finite-energy external force applied — can develop a singularity in finite time, an initially smooth fluid at rest eventually reaching unbounded speed (OpenAI; DataCamp). OpenAI states this establishes the Clay Institute’s official “statement C” (and also “statement D,” its periodic-domain counterpart) — one of four ways the Institute’s problem description allows the question to be resolved. This is distinct from statements A and B, which concern the unforced equations and which many mathematicians regard as the deeper, original question (DataCamp).

OpenAI’s own announcement describes the result as coming from an AI system “significantly more capable than GPT-5 Astra,” which coordinated roughly 10,000 concurrent agents working for about 88 hours, followed by roughly 17 more hours to formalize and machine-check the proof in the Lean proof assistant (OpenAI). A 166-page manuscript and the Lean formalization were published on GitHub the same day as the announcement (DataCamp). Clay Mathematics Institute president Martin Bridson called it “an exciting day, as we contemplate the announcement of major advances,” and mathematician Luis Martínez Zoroa of CUNEF University called it “a truly remarkable result” (Nature). OpenAI is explicit that it does not intend to claim the Millennium Prize itself for this result (OpenAI).

The announcement followed a burst of related work by other researchers. On September 7, 2026, mathematicians Levent Alpöge and Tristan Buckmaster of New York University published their own preprints, with public Lean formalizations, on related but distinct problems — finite-time blowup for the incompressible porous medium equation, the 2D Boussinesq system, and the 3D incompressible Euler equations (the zero-viscosity case) — using Anthropic’s Claude models together with OpenAI’s models. Separately that same day, Caltech researcher Anima Anandkumar released another solution using physics-informed neural networks (Nature; The Next Web). According to Axios, OpenAI CEO Sam Altman said the company’s own push began around September 1, 2026, after hearing rumors that Anthropic’s models had solved a major math problem, and that OpenAI wanted to test what its own system could do (Axios).

Verification status

As of this profile’s last update, independent mathematical peer review of OpenAI’s claim had not concluded. Reporting from the day of the announcement is not entirely consistent on timing: The Next Web wrote that OpenAI’s result was announced during a press call before the proof was available for outside scrutiny, quoting Buckmaster saying he “had not seen” it and noting that Buckmaster and Alpöge’s own preprints, by contrast, shipped with public, machine-checkable Lean formalizations from the outset (The Next Web). Other reporting the same day states the manuscript and Lean formalization were in fact published on GitHub on September 8 (DataCamp). This profile does not resolve that discrepancy in timing.

No source read for this profile documents a mathematician identifying a specific logical error or gap in OpenAI’s proof. Mathematician Luis Silvestre’s on-the-record reaction — “Yesterday and today are crazy days. We’re all, in the community, discussing the implications of this” — reflects the field’s shock rather than a stated technical objection (Scientific American). Terence Tao raised a different kind of concern, about process rather than correctness: he warned of “a substantial opportunity cost in converting a historically productive and motivating problem such as Navier-Stokes regularity into a mere viral social media post” (DataCamp), and separately warned that AI companies “strip-mining” open problems for solutions without preserving the reasoning behind them “may destroy the ecosystem” that produces future mathematical progress, since the dead ends along the way often carry the insight that helps mathematicians solve other problems (Fortune).

The credit dispute

Separately from the question of correctness, Buckmaster publicly disputed how OpenAI handled credit for the surrounding work. He said OpenAI researcher Sébastien Bubeck pressured him over authorship — proposing that Buckmaster publish jointly with OpenAI while being credited as closest to solving the problem, or publish alone and pursue the prize, but only if Alpöge’s name were removed because of his Anthropic affiliation. Buckmaster alleged Bubeck told him, “Why would you ruin your career?” Buckmaster also said he asked OpenAI directly whether its model had been trained on or had access to his and Alpöge’s private working sessions in Codex, and that he did not receive a clear answer (Fortune).

Bubeck denied the accusation, stating “We did not use their prompts or proofs to prompt our models,” and said OpenAI followed “academic norms,” later adding, “We recognize the priority of Levent Alpöge and Tristan Buckmaster’s work” (Fortune). OpenAI’s own announcement goes further in one respect, stating that it “cannot rule out that de-identified data derived from their usage of our products helped improve our models” (OpenAI).

What this profile does not claim

This profile does not assert that OpenAI’s proof is correct, complete, or accepted by the mathematical community — independent peer review had not concluded as of this profile’s last update, and OpenAI itself says it is not seeking the Millennium Prize for it. It does not treat OpenAI’s result as resolving the “harder,” unforced version of the Navier–Stokes problem (Clay’s statements A and B); it reports only what OpenAI itself says it proved, on the forced equations (statements C and D). It does not resolve the dispute between Buckmaster and OpenAI over whether OpenAI’s model had access, directly or indirectly, to Buckmaster and Alpöge’s prior work, nor the discrepancy between reports on whether the proof was available for scrutiny at the moment of announcement or only shortly after; both are presented here without adjudication. Sources also disagree on Levent Alpöge’s institutional affiliation (Harvard, per Nature, versus Anthropic, per OpenAI’s own post and other outlets), and this profile does not resolve that discrepancy. Finally, a September 10, 2026 report of separate, unconfirmed rumors that OpenAI or Anthropic were close to solving two different Millennium Prize problems (the Birch–Swinnerton–Dyer conjecture and the Hodge conjecture) is a distinct rumor thread from the Navier–Stokes claim described above; this profile does not treat those rumors as confirmed or as related to the events described here beyond both involving Millennium Prize problems.

Timeline

2026-09-01: OpenAI CEO Sam Altman says the company began its Navier–Stokes effort around this date, after hearing rumors that Anthropic's models had solved a major math problem and wanting to test its own system's capability.
2026-09-05: OpenAI states its internal AI system produced the result it would go on to announce, after roughly 88 hours of agent work plus about 17 hours to formalize the proof in Lean.
2026-09-07: Mathematicians Levent Alpöge and Tristan Buckmaster (NYU) publish their own preprints, with public Lean formalizations, on related but distinct problems: finite-time blowup for the incompressible porous medium equation, the 2D Boussinesq system, and the 3D incompressible Euler equations (the zero-viscosity case), using Anthropic's Claude and OpenAI's models. Separately that day, Caltech's Anima Anandkumar releases another solution using physics-informed neural networks.
2026-09-08: OpenAI publicly announces its claimed result on the forced Navier–Stokes equations, and publishes a 166-page manuscript and a Lean formalization on GitHub the same day. Clay Mathematics Institute president Martin Bridson calls it "an exciting day, as we contemplate the announcement of major advances."
2026-09-08: Buckmaster publicly disputes how OpenAI handled credit, saying OpenAI researcher Sébastien Bubeck pressured him over how Alpöge (because of his Anthropic affiliation) should be credited, and questioning whether OpenAI's model had access to their private working sessions. OpenAI denies using their prompts or proofs but says it "cannot rule out that de-identified data derived from their usage of our products helped improve our models." Separately, mathematician Terence Tao raises concerns about the opportunity cost of turning a long-studied open problem into what he calls "a mere viral social media post," without endorsing or disputing OpenAI's specific proof.

What we could not verify

These claims came up in research and could not be confirmed. They are recorded here rather than published as fact or quietly dropped. If you hold a source that settles one, please tell us.

OpenAI's claim concerns the "forced" branch of the Clay Mathematics Institute's official four-part problem statement (statements C and D), not the unforced branch (statements A and B) that many mathematicians regard as the original problem's harder, more central question. This profile does not treat resolving the forced branch as equivalent to resolving the unforced one; it reports what OpenAI itself says it proved.
Independent mathematical peer review of OpenAI's manuscript and Lean formalization had not concluded as of this profile's last update. Early reporting on the day of announcement (The Next Web) suggested the proof was announced before being made available for scrutiny, and quoted Buckmaster saying he "had not seen" it; other reporting the same day (DataCamp) says the manuscript and Lean files were in fact published on GitHub on September 8. This profile does not resolve that timing discrepancy.
No source read for this profile documents a specific mathematical error or logical gap in OpenAI's proof. The mathematician quoted expressing the most direct unease, Terence Tao, frames his concern as one of process and incentives (comparing the episode to a "viral social media post" and warning about "strip-mining" open problems for solutions without preserving the reasoning behind them), not as a claim that the proof is wrong.
Sources disagree on Levent Alpöge's affiliation: Nature's coverage describes him as a Harvard mathematician, while OpenAI's own post and other outlets describe him as affiliated with Anthropic. This profile does not resolve which description is more complete or current.
Whether OpenAI's model had any access to Buckmaster and Alpöge's private working sessions is disputed and unresolved: Buckmaster raised the question publicly and says he received no clear answer; OpenAI's Sébastien Bubeck denied using their prompts or proofs directly, while OpenAI's own announcement concedes it cannot rule out that de-identified usage data influenced its models. This profile states both positions without adjudicating between them.
A September 10, 2026 report described separate, unconfirmed social-media rumors that OpenAI or Anthropic were close to solving two different Millennium Prize problems (the Birch–Swinnerton–Dyer conjecture and the Hodge conjecture). These are a distinct rumor thread from the Navier–Stokes claim covered in this profile, remained unconfirmed by either company as of that report, and are not treated as resolved here.

Sources

8 sources (4 tier 1, 3 tier 2, 1 first-party), each opened and read against the claim it supports. Statements the subject makes about itself are marked as such and are not treated as independent evidence. See the source policy.

Built from Nature, Fortune, Axios, Scientific American, The Next Web, DataCamp and OpenAI's own announcement, all fetched and read in full. A first draft of this profile understated a key technical distinction (that OpenAI's claim concerns the forced equations, not the unforced formulation many consider the core of the problem) and omitted the separate verification-status dispute; both were added after a closer re-read of OpenAI's own material and additional independent coverage. A further, separate report (The News, Pakistan, September 10, 2026) about unconfirmed rumors of a different pair of Millennium Prize problems is cited only to show that those rumors are distinct from, and are not conflated with, the Navier–Stokes claim described here.

OpenAI claims huge maths breakthrough on a famed 'Millennium Problem' — Nature · Tier 1 Supports: The September 8, 2026 announcement itself, quotes from Clay Mathematics Institute president Martin Bridson and mathematician Luis Martínez Zoroa, OpenAI mathematician Sébastien Bubeck's quote, and the September 7, 2026 related solutions from Levent Alpöge/Tristan Buckmaster and Anima Anandkumar.
OpenAI says it cracked Navier-Stokes, one of math's grand challenges — Fortune, 2026-09-08 · Tier 1 Supports: Buckmaster's specific allegations against OpenAI and Sébastien Bubeck over credit, OpenAI's response and its acknowledgment of Alpöge and Buckmaster's priority, and Terence Tao's comment about AI "strip-mining" open problems.
OpenAI's historic math solution overshadowed by credit controversy — Axios, 2026-09-08 · Tier 1 Supports: Sam Altman's statement that OpenAI's effort began after hearing rumors of an Anthropic breakthrough, the scale of computing resources described, and further detail on the credit dispute with Buckmaster.
OpenAI claims blockbuster math breakthrough amid swirl of controversy — Scientific American, 2026-09-08 · Tier 1 Supports: Confirming that mathematician Luis Silvestre's on-the-record comment reflects the field's shock rather than a documented technical objection, and that no source read for this profile quotes a mathematician identifying a specific error in the proof.
OpenAI says it solved Navier-Stokes. Nobody has seen the proof. — The Next Web, 2026-09-08 · Tier 2 Supports: Early-day reporting that OpenAI's result was announced on a press call before the proof was available for outside scrutiny, and Buckmaster's statement that he had not seen it; also that Buckmaster and Alpöge's own preprints included public, machine-checkable Lean formalizations from the outset.
Did AI Solve Navier-Stokes? OpenAI's Claim, Explained — DataCamp · Tier 2 Supports: The precise technical framing (that OpenAI's proof targets the Clay Institute's official "forced" statements C and D, not the unforced A/B statements many consider the core question), the claim that a 166-page manuscript and Lean formalization were published on GitHub on September 8, and Terence Tao's separate comment about the "opportunity cost" of the episode.
AI rumors spark frenzy over two Millennium Prize Problems — The News International, 2026-09-10 · Tier 2 Supports: Documenting a separate, unconfirmed rumor thread (about the Birch–Swinnerton–Dyer conjecture and the Hodge conjecture) so it can be explicitly distinguished from, rather than conflated with, the Navier–Stokes claim described in this profile.
On the Navier–Stokes Millennium Prize Problem — OpenAI, 2026-09-08 · First-party Supports: OpenAI's own description of its method (the AI system, agent count, and runtime), its statement that the result concerns the forced equations and establishes the Clay Institute's statement C (and D), its explicit statement that it does not intend to claim the Millennium Prize, and its acknowledgment that it cannot rule out that de-identified usage data influenced its models.

Support independent research

Found this profile useful?

You can support FactsLedger by making a donation. Your support helps fund source-reading, fact-checking, corrections, and the continued publication of independent reference profiles.

Support the research