A public letter signed by 25 Fields Medalists is now criticizing rushed AI-generated mathematics, just days after OpenAI published a claimed solution to the Navier–Stokes problem. The headline is dramatic, but the status is more precise: OpenAI has published an analytic proof and a Lean formalization; independent mathematical acceptance remains pending.
That distinction matters. So do the separate questions around research credit and data use, where Tristan Buckmaster and Levent Alpöge’s related work has become part of a contested dispute.
What OpenAI claims
OpenAI says its internal AI system produced a construction in which a three-dimensional, incompressible Navier–Stokes flow with smooth external forcing develops a singularity in finite time while its energy remains finite. In plain English, the claimed mathematical flow becomes unbounded under the model’s equations.
The reported construction involves a vortex that spirals inward while axial stretching makes it increasingly elongated. OpenAI also says the result was formalized and checked in Lean, a proof assistant that mechanically verifies whether a formal proof follows from its encoded definitions and assumptions.
That is a published claim, not an officially accepted Clay solution. Navier–Stokes is one of the seven Millennium Prize Problems, each associated with a $1 million prize. OpenAI says it will not claim the prize for this result.
The animation helps make “finite-time blow-up” intuitive: it depicts a mathematical flow concentrating and stretching toward a singular behavior. It illustrates the construction OpenAI describes; it does not independently validate the proof.
Why Navier–Stokes matters
The Navier–Stokes equations describe how fluids move. The Clay problem asks whether smooth, three-dimensional, incompressible flow with constant density can remain smooth for all time—or whether a singularity can form in finite time.
A mathematical singularity is not a claim that real water or its molecules can travel infinitely fast. It is a behavior inside a continuous mathematical model: the velocity in the equations becomes unbounded under the stated conditions.
That distinction is easy to lose when “blow-up” is translated into everyday language. The result, if accepted in the relevant formulation, would address a deep question about the equations—not deliver an immediate upgrade to weather forecasting, aircraft design or other engineering systems.
What 10,000 agents actually tells us
OpenAI says approximately 10,000 concurrent agents worked on the effort for about 88 hours before reaching the reported result. GPT-6 Astra then spent approximately 17 additional hours formalizing and checking it in Lean.
For the Navier–Stokes effort, OpenAI reports about 2.7 million messages and approximately 130 billion output tokens. Those figures show the scale of the computation. They do not show that the system was more efficient than human mathematicians, nor do they establish that the proof is correct.
The resource use also changes how the achievement should be understood. This was not a short exchange in which a chatbot instantly answered a difficult textbook question. It was a large coordinated search followed by formalization. Whether that workflow becomes a reusable research method depends on what survives mathematical scrutiny.
Two research lines, not one result
The credit dispute is easier to follow when the mathematical targets are kept separate. Buckmaster and Alpöge were working on related forced-Euler and Boussinesq results with AI assistance. OpenAI says its published work concerns three-dimensional Navier–Stokes with smooth forcing and differs substantially from that research.
| Dimension | OpenAI’s reported result | Buckmaster–Alpöge related work |
| Mathematical focus | A claimed finite-time singularity for three-dimensional incompressible Navier–Stokes flow with smooth forcing | Related blow-up results involving forced Euler and Boussinesq equations |
| People and organizations | OpenAI researchers and an internal AI system | Tristan Buckmaster and Levent Alpöge, who is an Anthropic researcher |
| AI assistance described | Coordinated internal agents, followed by GPT-6 Astra for Lean formalization | Claude and Codex used as research assistants for documentation and logic steps |
| Relationship between the work | OpenAI says its result differs substantially from the related research | Buckmaster alleges that priority, credit and data-use discussions were mishandled |
Buckmaster alleges that OpenAI learned of the direction of his collaboration with Alpöge and handled credit discussions improperly. Those allegations remain contested. They should not be rewritten as a finding that OpenAI copied or misused the private work.
Direct access and possible model improvement are different claims
OpenAI denies that its researchers or agents saw Buckmaster and Alpöge’s specific work before it was publicly released. Separately, OpenAI says it cannot rule out the possibility that de-identified data derived from product use helped improve its models.
Those statements describe different pathways. The first concerns direct access during the proof effort. The second concerns possible indirect improvement of models from product-derived data. The second is not an admission that the researchers’ private work was used to produce the result.
The distinction may sound fussy. It is not. A claim about a model’s general improvement is not proof that a particular private proof or Codex session influenced a particular mathematical output.
This explainer organizes the chronology and discusses the technical and attribution dispute. Its commentary provides context; it is not independent mathematical acceptance of OpenAI’s claim.
Why Lean verification is not the same as acceptance
Lean can check that a formalized proof follows from the definitions and assumptions encoded in the system. That is valuable: formal checking can catch logical gaps inside the formal statement and implementation.
But a formal checker does not, by itself, settle whether the encoded statement matches the intended Clay problem. It also does not replace independent mathematicians examining the construction, its scope and its conceptual meaning. A proof can be mechanically consistent with a formalization while experts still debate whether the formalization captures the question that matters.
That is the central review task here. The existence of Lean code is an important artifact, not a substitute for broad mathematical understanding and acceptance.
The Fields Medalists’ warning goes beyond OpenAI
The open letter signed by 25 Fields Medalists argues that benchmark-driven AI mathematics can encourage rushed announcements, weak attribution and too little time for human understanding. Its concern is broader than this one dispute: mathematics is not only a sequence of formal steps, but also a body of ideas that researchers must be able to explain, examine and transmit.
That argument does not prove misconduct by OpenAI. It does identify the institutional pressure surrounding claims like this one. If a result is announced before its scope and lineage are clear, the public may receive a confidence signal before the mathematical community has had time to assess the substance.
For readers, the practical rule is simple: separate the artifact from the verdict. OpenAI has published a substantial computational and formalization effort. The verdict on the mathematical claim, the exact relationship to the Clay formulation, and the final allocation of credit still require expert scrutiny.
The next meaningful milestone is not another agent-count headline. It is whether mathematicians can independently examine the construction, agree on what problem it solves, and place its ideas—and its contributors—into the record accurately.
