OpenAI offered co-authorship if he dropped the Anthropic colleague

OpenAI says its agents solved the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems the Clay Mathematics Institute set in 2000. Only one of those seven had fallen before. The company ran about 10,000 agents at once on an internal model stronger than the Astra release, spent millions of dollars, and declined the $1 million prize.

The dispute that followed turns on authorship and access. MIT Technology Review laid out the sequence on Tuesday.

What happened, in order

• For roughly a year Tristan Buckmaster of NYU worked with Levent Alpöge, who is employed by Anthropic, using publicly available models from both OpenAI and Anthropic. They proved that a simplified form of the equations can break down.
• OpenAI learned of the effort. Sébastien Bubeck of OpenAI says the team was inspired to pursue the problem after hearing about it.
• OpenAI put two options to Buckmaster. Post first and OpenAI would publish the following day, or write a joint paper without Alpöge, because Alpöge works for Anthropic.
• Buckmaster posted his proof on Mastodon on Monday.
• He asked whether OpenAI's agents had read transcripts of his work, or whether models had been trained on them. Mark Chen, OpenAI's chief research officer, denied that any agent or employee accessed the transcripts. The training question went unanswered.

The condition is the story

Mathematics papers routinely carry authors from competing universities, competing countries and competing funders. A co-authorship offer conditioned on which lab a collaborator works for breaks with that.

Both proofs took the same route, an approach developed by Diego Córdoba and Luis Martínez-Zoroa. Javier Gómez-Serrano of Brown University said it was one of several thought to hold promise, which leaves independent arrival possible and influence unproven.

“Prematurely solving the problem by purely AI-powered methods can contaminate this process to the point where it actually becomes detrimental to mathematical progress.”

— Terence Tao, MIT Technology Review, 8 September 2026

Terence Tao, mathematician at UCLA, writing on Mastodon. Source: MIT Technology Review, 8 September 2026

What ten thousand agents cost

One side of this is two mathematicians and a year of work. The other is 10,000 agents running at once on an unreleased model, with a bill in the millions.

Tao is pointing at what falls between those columns. Mathematics has always moved on wrong turns, partial results and published mistakes, and those seed the next subfield. When an AI agent keeps its reasoning private, the field receives the result and none of the working.

We covered the same Astra family yesterday, when it finished Portal on its own, and this morning we looked at where AI stops helping in drug discovery. In both cases the machine cleared the mechanical step and the argument moved to the step where people judge.

The part nobody has answered

Whether OpenAI's models trained on Buckmaster's transcripts is a factual question with a yes or a no. Nobody at OpenAI has given one. Chen answered a narrower question, about access by agents and employees.

Until someone answers the wider one, the record shows two teams reaching the same result by the same route, one of them after hearing about the other.

This article is for informational purposes only and does not constitute investment advice.