OpenAI Says an Unreleased Model Solved a Navier-Stokes Millennium Prize Problem

On 8 September, OpenAI published a proof concerning the three dimensional incompressible Navier-Stokes equations, one of the seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000. The company says the work was produced by an unreleased model, described only as significantly more capable than GPT-6 Astra, running roughly 10,000 coordinated agents at once.

According to OpenAI's own account, the effort began on 1 September and reached a proposed resolution by 5 September, about 88 hours later. The output is a 166 page manuscript titled Finite Time Blowup for Navier-Stokes, alongside a formal verification of the argument written in the Lean proof assistant. The specific claim is narrower than solving the full Millennium Prize question: the system showed that an initially smooth fluid at rest can develop a singularity, a point where the equations break down, in finite time.

That distinction matters, because the Clay Mathematics Institute has not reviewed or endorsed the result, and a credit dispute broke out within days. Mathematicians who had been circulating earlier drafts and partial approaches to the same blowup question argued that the AI generated proof leaned on unpublished ideas from the wider research community without adequate acknowledgement. As of the most recent reporting, independent verification of the manuscript is still ongoing.

For anyone building with AI rather than proving theorems with it, the detail worth sitting with is the shape of the process, not the result. Ten thousand agents run in parallel for under four days, checked against a formal proof assistant that either accepts or rejects each logical step. That is a fundamentally different way of using a model than a single chat window, and it is the direction the frontier labs are now pointing their unreleased systems.

We build with AI at ZKO every week, mostly to pre visualise, draft, and iterate rather than to prove theorems. Watching a lab turn thousands of agents loose on a single unsolved problem, with a formal verifier as the only referee, is a useful reminder of how much headroom is still being worked out in how these systems get used, not just how large they get.