OpenAI says a newly trained internal AI model has resolved the Navier–Stokes existence and smoothness problem, one of mathematics’ seven Millennium Prize Problems, along with more than 100 other long-standing open problems across most areas of the field.
If verified, that would rank among the most consequential results produced by an AI system. OpenAI says the model’s progress surprised even its own mathematicians, prompting the company to establish an independent advisory group on mathematics and AI to help shape how such results are evaluated, communicated, and shared.
The announcement faces two kinds of scrutiny. Mathematicians must determine whether the proposed Navier–Stokes proof and the other claimed results survive detailed expert review. OpenAI must also establish credible processes for attribution, disclosure, verification, and responsible release when AI systems produce research faster than mathematicians can check it.
OpenAI’s Claim Extends Far Beyond One Famous Problem
OpenAI made the broadest version of its claim in a post on X, saying its internal model had resolved Navier–Stokes and more than 100 long-standing problems spanning much of mathematics.
The accompanying announcement is more measured. It says the internal model worked on a suite of open questions and produced several newly proven results. OpenAI has documented Navier–Stokes most extensively, but it has not yet published a complete catalog of the other problems, their previous status, or the proof artifacts supporting each one.
“Long-standing open problem” covers a wide range of mathematical difficulty. A problem could be famous and central to its field or highly specialized and known to only a small group of researchers. The age of a question alone says little about the depth of the proof required.
OpenAI therefore has two claims to establish. The Navier–Stokes result can be assessed through its public paper and supporting material. The claim of more than 100 resolutions requires a problem-by-problem accounting that identifies the exact statements, previous literature, AI contribution, human contribution, and current verification status.
The system appears to differ from a conventional chatbot producing a proof in one response. OpenAI describes a long-running research system composed of many coordinated AI agents. They can divide a problem into subgoals, investigate different proof strategies, use code, review intermediate work, and combine useful results into a larger argument.
For its Navier–Stokes experiment, OpenAI says the system maintained more than 10,000 concurrent agents over 88 hours. That scale resembles an automated research organization more than a single language model answering a prompt.
The Navier–Stokes Result Is the Clearest Evidence So Far
The three-dimensional Navier–Stokes equations describe how fluids such as water and air move. Given smooth initial conditions, mathematicians do not know whether the equations must continue producing smooth solutions indefinitely or whether a singularity can form in finite time.
A singularity would involve some part of the mathematical description becoming unbounded. In physical terms, quantities related to fluid velocity or its derivatives would grow without limit within a finite period. The Clay Mathematics Institute’s statement of the problem asks for a rigorous proof of global existence and smoothness, or a valid counterexample demonstrating finite-time breakdown.
OpenAI’s proposed solution takes the second route. Its Navier–Stokes research announcement describes a smooth, finite-energy initial configuration that allegedly develops a singularity in finite time.
The construction centers on a three-dimensional vortex with radial and axial flow. OpenAI says the model found a way to use the vortex’s geometry and symmetry to create a feedback process: axial stretching forces the vortex inward, inward contraction increases angular velocity, and the stronger rotation reinforces the stretching. Carefully chosen parity conditions and cancellations are intended to keep unwanted terms under control as the amplification continues.
According to OpenAI, the model produced the core construction, developed a working proof program, and connected arguments from several areas of mathematics. Dana Angluin, Lee-Ad Gottlieb, and Ben Lee Simpson ran the experiment, examined the output, and checked the proposed proof. The paper and supporting materials are available through OpenAI’s public supercritical-fluid repository.
A Lean formalization is also being developed alongside the conventional proof. Lean is a proof assistant that checks whether each formal step follows from specified definitions and axioms. A complete formalization could make certain classes of logical errors easier to detect, although formal verification does not automatically establish that the formal statement precisely matches the original mathematical problem.
OpenAI’s scientific capability evaluation protocol provides additional context. The company says it generated 3,000 independent outputs under a fixed experimental setup, with agents receiving a cached internet resource and standard code-execution tools. Only one run produced the full coherent approach that became the proposed solution.
If the proof is correct, the result is impressive. It also shows why the system should not yet be treated as a routine theorem-solving machine. The successful output was rare under the disclosed conditions, and expert selection remained essential. Researchers had to identify the promising run, reconstruct its argument, check the details, and prepare it for public examination.
A Claimed Solution Is Not a Millennium Prize Award
Calling Navier–Stokes “resolved” does not mean the Clay Mathematics Institute has awarded the associated $1 million prize.
The institute’s Millennium Prize rules set out a much slower process. A proposed solution must be published in a qualifying outlet, at least two years must pass, and the work must achieve general acceptance in the global mathematics community before the institute will consider awarding the prize.
OpenAI’s paper was announced on September 21, 2026. Regardless of the proof’s strength, it could not have completed that process on the day of publication.
Independent mathematicians must now examine whether every estimate holds, whether the constructed initial state meets all required conditions, whether the alleged blow-up occurs within the original Navier–Stokes system, and whether hidden assumptions entered the reduction. Specialists will also compare the construction with decades of previous work on vorticity growth, self-similar blow-up, and regularity criteria.
Lean could eventually provide another layer of confidence if the formalization covers the essential argument instead of selected lemmas. The proof assistant’s foundations, imported libraries, numerical assumptions, and correspondence with the paper would all need to be clear.
Such scrutiny is normal mathematical practice. Human-authored proofs of major results can spend years under review, and apparent breakthroughs have collapsed over subtle errors. A Millennium Prize claim demands at least the same standard.
OpenAI Is Building an External Review Layer
OpenAI’s new advisory group seeks to address the institutional questions accompanying the mathematical ones. Its initial members are:
- Susan Athey of Stanford University
- Kevin Buzzard of Imperial College London
- Chandler Davis of the University of Toronto
- Kristin E. Lauter of Meta AI Research and the University of Washington
- Terence Tao of the University of California, Los Angeles
- Gero von Randow of Die Zeit
The group plans to advise OpenAI on frameworks for assessing and communicating AI-generated mathematics, maintaining academic and professional standards, and building tools for mathematical research and education. OpenAI says it wants mathematicians involved in deciding how AI supports mathematical understanding and how its benefits reach the wider community.
The initiative follows an August 2026 open letter on AI and mathematics calling for independent structures to guide the field. The letter emphasizes human mathematical culture, transparent professional standards, responsible evaluation, and broad access to the benefits of AI-assisted research.
OpenAI will provide administrative funding but says it will not determine the group’s activities or recommendations. Members generally do not receive compensation from the company, although von Randow holds a paid administrative role funded through the group.
The advisory group is not a peer-review panel for OpenAI’s proofs. Its members will not receive confidential access to OpenAI’s systems, conduct technical assessments of individual model outputs, or decide whether a particular result should be released.
Its formation should therefore not be presented as independent confirmation of the Navier–Stokes solution. The group will focus on governance, standards, communication, and tool development. Specialists studying the actual mathematics must still validate the proof.
The 100-Problem Claim Needs a Public Research Ledger
If OpenAI’s internal model is resolving open problems at the claimed pace, conventional academic publishing will struggle to keep up. Even correct proofs require experts to read them, reconstruct their reasoning, compare them with existing literature, and rewrite them in forms other mathematicians can use.
A public research ledger would make the broader claim easier to evaluate. For each problem, OpenAI could disclose:
- The exact mathematical statement and field
- When and where the problem was previously posed
Frequently Asked Questions
5 questions
1Did OpenAI’s AI officially win the Navier–Stokes Millennium Prize?
No. OpenAI announced a proposed solution, but the Clay Mathematics Institute has not awarded the prize. Its rules require publication in a qualifying outlet, a waiting period of at least two years, and general acceptance by the global mathematics community.
2
Sources
- independent advisory group on mathematics and AIopenai.com
- https://x.com/OpenAI/status/2102093145051943229x.com
- Clay Mathematics Institute’s statement of the problemclaymath.org
- Navier–Stokes research announcementopenai.com
- public supercritical-fluid repositorygithub.com
- scientific capability evaluation protocolopenai.com
- Millennium Prize rulesclaymath.org
