OpenAI Claims a Navier–Stokes Solution: What Is Known

On September 8, OpenAI announced a solution to the Navier–Stokes problem.

The equations describe liquid and gas flow. This is one of Clay’s seven Millennium Prize Problems, each allocated a $1 million prize. That is the prize amount, not an announcement that it has been awarded.

What the proposed result says

The OpenAI paper describes a three-dimensional flow starting from rest under a smooth external force. Its authors claim that velocity becomes unbounded in finite time while kinetic energy stays bounded. This mathematical breakdown is called a singularity. It is a construction within the equations: water in a physical experiment was not accelerated to an infinite speed.

In the authors’ account, a central vortex contracts and speeds up. The difficult part is showing that the applied force remains smooth: the unbounded growth must arise from the flow’s dynamics, rather than being inserted into the force itself.

The Clay Mathematics Institute’s problem statement offers four alternatives. Versions A and B concern smooth solutions without an external force; C and D allow a breakdown example with smooth forcing. The new paper targets C and D. External forcing therefore does not, by itself, violate the prize’s stated conditions. Equally, this result does not establish breakdown in the unforced case.

Model roles

OpenAI says an internal model more capable than GPT‑6 Astra searched for the proof; Astra performed Lean formalization and verification.

For background on dividing work between software workers, see our explanation of AI agents.

What Lean verification adds

Lean checks logical derivations from definitions, earlier theorems and axioms, independently of how convincing a model’s prose sounds. However, the Lean documentation separates proof validity from the meaning of the theorem being proved. Reviewers must also check whether the formal statement represents the original mathematical question. A successful check of a different statement would not answer that question.

OpenAI has published the formalizations in a repository, with build instructions and a route to independent checking. Specialists can examine those materials. RozumTech has not independently audited the mathematical proof.

Has the problem officially been recognized as solved?

When checked on September 9, the Clay Mathematics Institute’s page still listed the problem as unsolved. That records the page’s status; it is not a finding by the institute that the proposed proof is wrong.

OpenAI says it does not intend to claim the prize.

Before considering an award, Clay’s rules require publication in a qualifying outlet, at least two years since publication and general acceptance by the global mathematics community. That waiting period governs the prize procedure; mathematicians can examine the work now.

For readers, the key development is whether specialists confirm the stated result. A computational achievement at this level deserves attention, without implying that chatbots now solve every scientific problem without error.