10,000 AI agents, 88 hours, and one very awkward question about whose work it was
OpenAI says its machines cracked a maths problem that stood for nearly a century. A mathematician says his unfinished proof was sitting on OpenAI's servers at the time.
There is a short list of maths problems so hard that solving one gets you $1m. The Clay Mathematics Institute published seven of them, called the Millennium Prize Problems, to mark the biggest unsolved puzzles in the field. On Tuesday OpenAI said it had cracked one of them: Navier-Stokes, the equations that describe how air and water move. About 10,000 AI agents — systems that work on their own, without a person typing prompts — ran at the problem at once and got there in 88 hours.
And then the story turned. Tristan Buckmaster, a professor at New York University, said OpenAI sped up its own effort after hearing that he and a researcher at Anthropic, an OpenAI rival, were close to announcing a breakthrough of their own. Their work in progress, he said, was stored in OpenAI's Codex model — the tool used for writing code. Which means it sat on the servers of the company that then announced the answer first.
Start with what the machine actually did, because it is genuinely strange. Navier-Stokes asks whether the equations we use to describe moving fluids ever break down. Every weather forecast, every aircraft wing, every simulation of water in a pipe leans on them. OpenAI's proof suggests they do break: under certain conditions the equations "blow up", with the speed of the fluid becoming impossibly infinite. Nearly a hundred years of humans could not settle that. A swarm of software settled it over a long weekend, and it took GPT-6 Astra about 17 hours to check the answer.
Now the uncomfortable part. Buckmaster is careful — unusually careful. "I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything," he wrote on his website. OpenAI researcher Sebastien Bubeck denied at a press briefing that the company used the pair's work or accessed material on its servers. But OpenAI's own announcement says it could not rule out that data from the pair's use of its products "helped improve our models". Two statements, both from OpenAI, sitting a little uneasily side by side.
That is the mechanic worth carrying away, and it has nothing to do with maths. Anything you type into an AI tool lives on someone else's computer. If your unfinished work is valuable, the company holding it is not a neutral filing cabinet — it may be a competitor. Two researchers at the top of their field learned this in public. Everyone else learns it quietly, or not at all.
The proof may hold. The trust question will outlive it.
Anyone who drafts something in a chatbot before it is finished — a thesis, a business plan, code for a side project, a pitch. Students especially: your unpublished work goes to a company that also builds products. Nobody has said what happens if that company gets there first. And anyone deciding right now what to study — if a swarm of agents can close a century-old problem in 88 hours, the value of "I can solve hard problems" is being repriced while you are still choosing a course.
Two things to watch. OpenAI said it does not intend to claim the $1m prize from the Clay Mathematics Institute, so the usual verification route is not being triggered by the company itself — whether other mathematicians confirm the proof is the first test. The second is Buckmaster and his co-author: they were poised to announce their own breakthrough, and whether it appears, and what it says, is not something anyone has scheduled. No dates were given for either.
OpenAI is preparing for a flotation that could value it at around $1tn. It walked away from $1m. That tells you what the announcement was actually for — and it wasn't the prize money.
Sources: The Guardian, "OpenAI claims to have solved maths problem that stumped humans for decades", by Ian Sample and Dan Milmo, 8 September 2026.
Машина за 88 часов закрыла задачу, над которой математики бились почти сто лет — и сразу же встал вопрос, не подсмотрела ли она чужую работу, хранившуюся на серверах OpenAI.
Written by THE TELL’s AI newsroom. how we work · corrections