OpenAI claims 10,000 agents solved the 90-year-old Navier-Stokes problem

OpenAI announced on Tuesday that it has found a solution to the Navier-Stokes problem, a question about the behaviour of flowing liquids and gases that has stood open for roughly 90 years, according to The Verge. The company credited an internal AI model it says is more capable than its newly released GPT-6 Astra, working alongside about 10,000 agents running concurrently.

The Verge notes the claim was first reported by The New York Times and Wired, with OpenAI following up in a blog post. Navier-Stokes is one of seven Millennium Prize Problems, each carrying a $1 million reward, and OpenAI says it does not plan to take the prize money.

Ten thousand agents and millions of dollars

According to Axios reporting cited in a dev.to analysis, OpenAI said an internal model group began the effort on September 1, after earlier solutions had been reported, and that the resulting proof points to a finite-time singularity in the three-dimensional equations. The run reportedly lasted about 88 hours and cost millions of dollars. The Verge separately reports that OpenAI began training the internal model on August 28 and that it delivered standout results on the company's own benchmarks, mathematics included.

The scale is the notable part. Rather than a single model producing a single answer, the account describes thousands of agents working in parallel, which suggests dividing the problem, generating candidate arguments and cross-checking intermediate steps. As dev.to points out, however, nothing public yet explains how those agents were assigned, supervised or evaluated.

A dispute over data and credit

One day before OpenAI's announcement, New York University mathematics professor Tristan Buckmaster published findings on a related problem together with Levent Alpöge, a researcher at Anthropic. According to The Verge, Buckmaster says that after learning OpenAI had heard about their progress, he contacted the company and found it had produced a proof along a route he and Alpöge had been exploring, using OpenAI's Codex and Anthropic's Claude, with all of their project drafts stored in Codex sessions.

Buckmaster says he asked whether the model had access to, or was trained on, those sessions. He says he was told the model does not look up user data, but that he received no answer when he asked again about training.

OpenAI, in a statement to The Verge that echoes its blog post, said no specific user data was accessed in order to solve the problem, while acknowledging it cannot rule out that de-identified data derived from use of its products helped improve its models. OpenAI's Sebastien Bubeck added that the company saw none of the two researchers' work before its public release, and that the proofs differ considerably, including in the precise results proved. Buckmaster responded on Mastodon that the statement amounts to OpenAI admitting it used training data from a period after he and Alpöge had found their result.

What is and isn't verified

A reported proof is not an accepted solution. Dev.to stresses that there is no publicly available technical paper, preprint or independently inspectable argument behind the claim, and that the mathematical community would need access to the full proof plus a process of detailed checking before treating it as settled. The accounts also diverge on publication status: The Verge describes a Tuesday blog post from OpenAI, while the same-day dev.to analysis says it could find no accessible first-party technical write-up, only media reporting of the result.

Why it matters

If the proof holds up, this would be an AI-assisted crack at one of mathematics' hardest open questions, and evidence that large coordinated agent systems can take on work requiring sustained reasoning, iteration and verification rather than one-shot answers. The credit dispute matters just as much: researchers who use AI coding tools now have to weigh whether their drafts could indirectly sharpen a lab's own models, and OpenAI's own wording leaves that possibility open. Until a complete proof is published and independently checked, this remains a landmark claim rather than a landmark result.