OpenAI’s Navier–Stokes Proof Claim: What Is Verified?

*Source refresh: October 21, 2026. DiscoverAI has not independently checked the mathematics or reproduced the Lean build. This page reports a published claim and explains the verification boundary.*

What did OpenAI announce?

OpenAI's September 8 announcement says an internal AI system produced an analytical proof and Lean formalization for a forced three-dimensional Navier–Stokes singularity. It identifies alternatives C and D in the official formulation and links the writeup and formal repository. The company says the internal model was more capable than GPT-6 Astra.

Those are attributed claims and inspectable artifacts. They should not be rewritten as a confirmed prize award or universal solution to every fluid-dynamics problem. The Clay Mathematics Institute's problem page and official formulation remain essential references for interpreting the target theorem.

Three different verification questions

A positive answer to the first question does not answer the other two. Even a successful formal build needs an assessment of whether the formal statement matches the intended mathematical problem.

What Lean can establish

A proof assistant checks that encoded conclusions follow from encoded definitions, assumptions and accepted dependencies. That can substantially reduce certain classes of reasoning error. It does not automatically tell a general reader whether the theorem has the exact scope suggested by a headline.

A responsible reproduction record should identify repository commit, toolchain version, dependency versions, build instructions and unresolved placeholders or extra assumptions. It should distinguish a clean independent build from a screenshot or a statement that someone else ran it.

Why scope matters

Forced and unforced equations are different questions. Smoothness assumptions, domain, initial conditions and the permitted external force all matter. Do not paraphrase a carefully bounded construction as a claim about all real fluids or all numerical simulations.

The original formulation should be read alongside any proposed proof. A commentator's intuitive description is useful for orientation but cannot replace the theorem statement. Ask which alternative is addressed and what remains outside the result.

What evidence would change this page?

A useful update would include an independently reproduced formal build, a named expert analysis of the exact theorem, an author correction or an official institutional assessment. Each should have a date and a source. A social post celebrating a breakthrough is weaker evidence than a reproducible artifact or substantive mathematical review.

This refresh does not establish a prize award or independent mathematical consensus. Avoid interpreting the absence of such evidence on this page as a proof that the result is false; it identifies the limits of this editorial review.

What to learn about AI research claims

Judge the inspectable result separately from compute scale or model branding. Retain the problem statement, artifact version and verification method. For practical research workflows, apply the same discipline to citations: a linked source should support the exact claim made, and an unsupported conclusion should remain unresolved.

The 25-question evidence benchmark provides a smaller operational example of recording sources, contradictions and unknowns. It does not validate this mathematical result.