Of all the AI news this month, this is the one that made me sit up. OpenAI has claimed that a swarm of AI agents produced a proof addressing the Navier-Stokes existence and smoothness problem — one of the seven Millennium Prize Problems, each carrying a $1 million bounty from the Clay Mathematics Institute. And the fresh twist this week is a bitter fight over who deserves credit.
The claim, in plain English
The Navier-Stokes equations date back to the 1840s, and they describe how fluids move — water through pipes, air over an aircraft wing, blood flow, weather systems. Engineers use them daily. But mathematicians still can't answer a basic question about them: if you start with perfectly smooth conditions, do the equations always stay smooth forever? Or can they "blow up" — predict a fluid moving infinitely fast, something no real fluid can do?
OpenAI's proof shows the blow-up scenario is real — under specific conditions. With a carefully designed, smooth external force applied to the fluid, the equations can break down and produce a singularity in finite time. That settles the "forced" versions of the problem (alternatives C and D in the Clay Institute's official statement, which explicitly allow this route). The harder, unforced version — blow-up with no external push at all — remains open.
How 10,000 agents spent 88 hours on one math problem
According to OpenAI's account, the effort began September 1, after researchers heard rumors that others were closing in on Millennium problems. Agents were organized into groups that could read a cached snapshot of the web, run code, and message each other. A smaller group of roughly 100 agents cracked a related Euler equations question in about 50 hours. Then the effort scaled up — to around 10,000 concurrent agents working on Navier-Stokes. They reached their conclusion on September 5, about 88 hours in, exchanging 2.7 million messages and generating around 130 billion output tokens along the way.
Then came the verification. GPT-6 Astra spent another 17 hours translating the argument into Lean, a programming language where computers check every line of a proof — meaning the logic can be machine-verified rather than taken on trust. The final paper runs 166 pages. The underlying model? Undisclosed, except that OpenAI says it is "significantly more capable than GPT-6 Astra" (which we covered in detail earlier) and was built with large-scale reinforcement learning. Outside estimates of the compute bill range from $2 million to $22.5 million at list prices.
The credit fight nobody saw coming
Honestly, what surprised me most isn't the proof — it's the drama around it. Tristan Buckmaster of NYU and Levent Alpöge, a mathematician who works at Anthropic, had spent nearly a year pursuing the same approach, using OpenAI's Codex and Anthropic's Claude as research tools. On August 15 they proved that the Euler equations can blow up, a result Fields Medalist Terence Tao called "a remarkable achievement." On September 7, they posted Lean-verified proofs of three related fluid equations. OpenAI published its paper on September 8 — one day later.
Buckmaster publicly accused OpenAI's math lead, Sébastien Bubeck, of pressuring him over credit for the earlier work, and raised the uncomfortable question of whether OpenAI's model had absorbed ideas from his private Codex conversations. OpenAI's response evolved: first it said it couldn't determine whether the pair's activity had influenced its models, then, after an internal investigation, it stated that Buckmaster's Codex prompts in the two months before the announcement "could not have influenced the system in any way, including through training." It hasn't addressed earlier prompts. A deep-dive published by DeepLearning.AI's The Batch on Friday walks through the whole dispute — and notes that 25 living Fields Medalists have now signed a declaration criticizing AI labs for claiming credit on famous open problems.
So is the problem solved?
Not officially — and that's worth being clear about. The Clay Mathematics Institute acknowledged the apparent resolution of the forced cases but said its evaluation will be deliberately unhurried. Its rules require publication, a two-year wait, and broad acceptance by expert mathematicians before any prize is considered. OpenAI, for its part, says it isn't seeking the million dollars. The Lean formalization means the logical checks can be re-run by anyone, which is genuinely new — but "it compiles" is not the same as settled mathematics.
Why this matters beyond mathematics
Buckmaster's own framing might be the most quotable line of the whole affair: a mathematician working with an AI model can now do this kind of work in a month. That's the real signal — not that mathematicians are obsolete, but that the throughput of AI-assisted research just jumped by an order of magnitude. It also puts a spotlight on agent costs: 130 billion tokens for one problem is real money, and if you're experimenting with multi-agent setups yourself, cost discipline is a survival skill. Our GPT-6 Astra — Zero to Hero Guide includes a full chapter on cost optimization (plus one on building your own AI agents) for exactly this reason. And if AI-versus-math stories are your thing, GPT-6 Astra was also recently credited with the core idea behind the first-ever "Major Advance" problem solved on Epoch AI's FrontierMath — we covered that one too.
Key Takeaways
- OpenAI's unreleased model produced a proof of finite-time blow-up for the forced Navier-Stokes equations — an officially allowed route to the $1M Millennium Prize problem
- ~10,000 concurrent agents, 88 hours, 2.7 million messages, ~130 billion output tokens; verification in Lean took another 17 hours
- The unforced version of the problem remains open, and Clay's review process is deliberately slow
- Buckmaster and Alpöge had the closely related Euler result first; OpenAI denies their work influenced its model
- The bigger signal: months of expert mathematical work compressed into days by coordinated AI agents
Frequently Asked Questions
What does OpenAI claim its agents did?
OpenAI says an unreleased internal model — "significantly more capable than GPT-6 Astra" — produced a proof of the forced Navier-Stokes problem, one of the seven Millennium Prize problems worth $1 million each.
How did the agents work on it?
About 10,000 concurrent AI agents worked on it, exchanging 2.7 million messages over 88 hours, with verification in Lean taking another 17 hours.
What is the controversy around the proof?
Two mathematicians — Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) — had been chasing the same result for a year and published one day before OpenAI. OpenAI says it won't claim the $1M prize.
What This Means for You
If you're a student in India, this should settle the old "maths or coding?" debate for good — the answer is now maths and AI. The tools that solved a 180-year-old fluid dynamics question are the same class of tools you can learn to direct today, and the gap between "knowing how to prompt" and "coordinating a swarm of agents" is where the interesting careers are forming. You don't need $22 million of compute to start; you need to understand how these systems think, how to keep their costs under control, and how to verify what they produce. That's exactly the path our GPT-6 Astra — Zero to Hero Guide walks you through, from your first serious prompt to running your own agent workflows.