OpenAI logo with the OpenAI emblem Credit: Evolf via Shutterstock.com When OpenAI announced on a press call that an internal model had cracked one of mathematics’ hardest open problems, we wrote that nobody had seen the proof. That is no longer true, and what has now been published narrows the claim in ways the announcement did not.OpenAI has posted a write-up, a PDF paper, and a formalisation of the argument in Lean with a public GitHub repository. The formalisation is the part that matters most, because a Lean proof can be machine-checked by anyone, which removes the need to take the company’s word for the mathematics. OpenAI says the formalisation and verification took a further 17 hours using GPT-6 Astra.
The result itself is that an initially smooth fluid at rest can develop a singularity in finite time. In OpenAI’s framing, this “resolves the Navier-Stokes Millennium Prize problem by establishing statement ‘C’ (and also ‘D’) in the official Millennium Prize formulation”. And then, two paragraphs later: “We do not intend to claim the Millennium Prize for this result.” That sentence does more work than the announcement did. A company that had unambiguously solved a problem carrying a $1M prize and 26 years of failed attempts would claim it, and would enjoy doing so.
The decision not to is the clearest available signal about how OpenAI itself rates the strength of the claim against the prize’s formal criteria. The crux is forcing. OpenAI’s fluid has “a smooth force applied to it” throughout, from rest to the formation of the singularity. The company draws the contrast itself when describing separate work on the Euler equations by Levent Alpöge and Tristan Buckmaster, noting that the variant they resolved “was the unforced version, where no external force is applied”.
Whether a blowup produced under an applied force satisfies the Clay Institute’s formulation is the question everything else rests on, and it is a question for mathematicians rather than for a technology desk. The scale of the attempt is its own story. Roughly 10,000 concurrent agents worked the problem across about 88 hours, from 1 to 5 September, consuming 2.7 million messages and around 130 billion output tokens. The model was not GPT-6 Astra, which shipped last week, but an internal system OpenAI describes as significantly more capable.
Astra was relegated to checking the answer. There is a broader shift buried in the method. The proof was not produced by one model reasoning at length, but by a swarm of them running in parallel for three and a half days, at a compute cost OpenAI has previously put in the millions. That is a different research instrument from the one the field has been benchmarking, and it suggests the frontier is moving from how clever a single model is to how much of it you can afford to point at one question.
It also means results of this kind will be reproducible only by organisations that can spend at the same scale, which is an uncomfortable property for mathematics to acquire. The credit dispute has moved too, and in an interesting direction. When we reported the announcement, Buckmaster had questioned whether OpenAI pursued directions derived from his unpublished work with Alpöge, and OpenAI denied it. The published page now credits both men for concurrent work on the forced Euler problem and offers to recognise their priority in a joint announcement.
A company that considered the allegation baseless does not usually respond with a priority offer within days. What is now verifiable and what is not have separated cleanly. The mathematics can be checked by anyone willing to run the Lean proof. Whether the result meets the Millennium criteria is a matter for the Clay Institute and the field.
Where the ideas came from is the one part that no amount of formalisation can settle, because provenance is a claim about a training process nobody outside the company can inspect.
