OpenAI announced that its unreleased internal model, significantly more powerful than GPT-6 Astra, obtained a proof of finite-time 'blowup' of smooth solutions to the 3D Navier-Stokes equations—one of the seven Clay Institute Millennium Prize Problems. The statement turned into a conflict with the mathematical community: 25 Fields Medalists signed the declaration 'A Severe Misalignment of AI in Mathematics,' and OpenAI responded by opening an archive of 722 machine-generated manuscripts. There is currently no independent verification of the claimed proof.

image
image

What Happened

In August 2026, OpenAI gathered about 40 mathematicians to discuss how its internal models are already handling open research problems. On September 8, the company reported that an unreleased model constructed a proof of finite-time 'blowup' of smooth solutions to the 3D Navier-Stokes equations: approximately 88 hours of work by roughly 10,000 parallel agents, about 130 billion output tokens per run, and formalization of the result in the Lean language. A scandal then erupted: NYU mathematician Tristan Buckmaster and Levent Alpöge, working at Anthropic, stated that OpenAI had 'intercepted' their joint work and had offered Buckmaster to write a paper without Alpöge because the latter works for a competitor. On September 11, 25 Fields Medalists, including Terence Tao, Peter Scholze, James Maynard, and Maryna Viazovska, signed the declaration 'A Severe Misalignment of AI in Mathematics.' On October 6, OpenAI opened the repository openai/math: it contains 722 manuscripts with reasoning traces and Lean formalizations, grouped into 372 'families' of results.

Context

The Navier-Stokes equations describe the flow of viscous fluids and gases, and the question of smooth solutions—whether fluid motion can 'blow up' to infinity in finite time—has been listed among the seven Clay Institute Millennium Prize Problems since 2000. Mathematicians have worked on them for years, relying on accumulated partial results, so a claim of a complete proof would be a sensation in itself. The resonance was amplified by the publication format: the result was not released through a peer-reviewed journal, but as a company statement accompanied by a dump of artifacts on GitHub—in the middle of a race for priority, where such formats break the traditional chain of attribution, citation, and inclusion of ideas into the scientific canon. Hence the wording of the declaration: the dispute is not about whether the model is good, but about who controls the research process—several laboratories with large computational resources or the open scientific community.

Why This Matters for the Industry

For the industry, this is primarily an engineering rather than an academic signal: it demonstrates a production template for intellectual labor—a swarm of parallel agents over a long horizon, formal verification, and publication of raw materials, manuscripts with reasoning traces, instead of peer-reviewed articles. Solving fundamental problems is shifting from 'talent plus years' to 'computations plus days,' and what becomes defensible for companies is not 'we are smarter,' but access to computations, a layer of independent verification, and the ability to embed machine-checkable conditions into their pipeline. There is no product that can be deployed to production today: the model is closed, and API, pricing, and latency metrics have not been announced, so the deployable component is the method, not a release. The practical move right now is to dissect the open repository openai/math: the structure of the manuscripts, reasoning traces, and methods for organizing and filtering thousands of agents, as well as prototyping navigation and explanation over such a corpus.

Why This Matters for Users

For the reader, this is a live demonstration that AI is changing science right now, not in some distant future. The practical step is simple: read the declaration by the 25 laureates on mathandai.org and Terence Tao's post on 'Math 2.0,' then open the GitHub repository openai/math and see how the machine-generated manuscripts are structured—there are reasoning summaries on the irrational power of the number π, on Mahler's conjectures, and on the Kaplansky constant. The main shift visible from these materials is that the task has shifted from 'find the answer' to 'explain and verify why this is the answer,' and the formula 'verifiable does not mean understandable' takes on a literal technical meaning. This experience also forms a useful habit: to treat strong AI claims as hypotheses and to ask companies not for a press release, but for access to artifacts and the status of independent verification.

What Is Still Unknown / Limitations

The claimed result is currently unverified: there is no independent confirmation of either the Lean formalization or its correspondence to the arXiv:2610.08144 preprint—it is known that the Lean version of the proof does not match the text of the article, meaning that machine checking does not yet guarantee the correctness of this specific result. The selection effect has not been measured: it is unknown how many attempts and failures stand behind the 372 successful 'families,' so the success rate remains unknown, and the corpus of manuscripts cannot be read as an assessment of the model's overall capability. Reproducibility for external parties is zero: the model has not been released, there is no API, and it is impossible to replicate the run budget from the outside. The public part of the conflict with Buckmaster and Alpöge relies on the statements of the participants themselves, and until documents from both sides appear, it should be described cautiously.

Sources

Author

Look at AI, editorial team