An NYU mathematician says OpenAI raced him to a Millennium Prize problem

He does not believe the approach was arrived at independently in the time available: "It is not the direction one arrives at in a few days by giving a model the problem statement."

The framing of his complaint is worth quoting, because it is not really about priority. "There is another part of this story, and one that, honestly, I very much wish I did not have to be concerned with."

OpenAI's response is a direct denial: "We (the researchers and the agents) did not see any of their work through any means until they released it publicly."

The scale of what OpenAI ran

OpenAI's effort began on 1 September, prompted by rumours that Millennium Prize problems were being solved, and used an unreleased next-generation model. It consumed 300 billion output tokens, which the company puts at about $22.5 million in compute.

That figure is the most informative number in the story, and it is the one that should be repeated whenever a lab announces a mathematical result. Twenty-two and a half million dollars of inference on a single problem is not a demonstration of efficient reasoning. It is a demonstration that a sufficiently funded organisation can brute-force a search that a mathematician would conduct over years.

It also sets a price on this kind of result that only a handful of organisations can pay.

The contrast with the week's other mathematical result

Anthropic announced on 4 September that Claude had produced a fully computer-checked formalisation of Fermat's Last Theorem in Lean, translating an existing proof over 11 days.

That result makes a smaller claim and rests on stronger evidence. Formalisation is mechanically verified — Lean accepts the proof or it does not — and translating a known proof does not assert mathematical discovery.

A claimed novel proof of a Millennium Prize problem is the opposite: an enormous claim whose verification depends on human referees, over months, and which has not happened. OpenAI has published a proof. Whether it is correct is not yet established by anyone.

What the dispute is actually about

Strip out the priority question and what remains is a governance one. A research programme with $22.5 million of discretionary compute, an unreleased model, and no publication norms can enter any open problem in mathematics at a moment of its choosing.

Mathematics has customs for this — people announce what they are working on, and others give them room. Those customs are enforced by reputation within a small community, and they have no purchase on an organisation outside it.

The Clay Mathematics Institute has not commented on whether the proof will be considered for the prize, which requires publication in a refereed journal and a two-year waiting period. Neither has been satisfied.