OpenAI says it solved a $1m maths problem. Here's what that actually means

You may have seen headlines saying AI 'solved' a famous $1 million maths problem in September 2026. It's a genuinely interesting story — just not quite the clean, total victory the headlines implied. Here's what actually happened, in plain English.

Think of any fluid — water, honey, air — as something you can describe with maths rather than by tracking every single molecule. The Navier–Stokes equations do exactly that: they're essentially Newton's famous 'force equals mass times acceleration' rule, applied to a moving fluid. They're used to design aircraft, forecast the weather, and model blood flow. The trouble is a deep mathematical question nobody could answer: if you start with a perfectly smooth, calm fluid, could the equations ever predict that some tiny point inside it starts moving infinitely fast, in a finite amount of time? That's what mathematicians call a 'blow-up' or a 'singularity'. It can't happen to a real fluid — real fluids are made of atoms, and atoms can't move infinitely fast — so a blow-up would mean the equations themselves have a hidden breaking point. In 2000, the Clay Mathematics Institute put a $1 million prize on settling this, one of seven famous 'Millennium Prize Problems'. Only one of those seven — the Poincaré Conjecture — has ever been officially declared solved.

OpenAI says it pointed a powerful, unreleased internal AI model — more capable than its public GPT‑6 Astra model — at the problem, using roughly 10,000 AI 'agents' working together and talking to each other (nearly 5 million messages, across everything they attempted). After 88 hours, that group found a solution: a spinning vortex that squeezes inward and speeds up.

Picture stirring a cup of tea faster and faster at the very centre, while the swirl there gets thinner and thinner, like a strand of spaghetti being pulled tight — until, mathematically, it reaches infinite speed. A further 17 hours went into checking the proof using something called Lean — think of it as an incredibly strict, automatic examiner that reads every single logical step and simply won't accept the proof unless each one is airtight. That's an important safeguard, because AI can sound completely confident while being wrong.

Here's the bit that most headlines skipped. OpenAI's fluid didn't break down entirely on its own — it needed a carefully engineered outside push, a bit like gravity or a fan blowing on it, applied in just the right way. The Clay Institute's own official rules do allow this as a valid way to win (it's one of the accepted 'versions' of the question), so nothing underhand happened. But most mathematicians have always cared more about the harder version: can a fluid go haywire with no outside push at all, purely from its own churning? That version is still unsolved, and a proof released on 17 September 2026 shows the method OpenAI used can't be stretched to reach it.

No — OpenAI has said plainly it doesn't intend to claim the prize. Even if it wanted to, the Clay Institute's rules require the proof to be published properly, survive a minimum two-year waiting period, and win broad acceptance from the maths community first. That process hasn't even started. Martin Bridson, the Institute's president, called the announcement exciting without declaring anything settled.

There's also a human side worth knowing about. About 12 hours before OpenAI's announcement, a New York University mathematician named Tristan Buckmaster, working with an Anthropic researcher, put out a related (but different) proof of his own, using several AI tools together. Charles Fefferman — the Princeton mathematician who originally wrote this problem's official description in 2000 — said he was thrilled the underlying maths problem had finally moved, and was quick to credit the two researchers whose years of patient, non-AI work made all of this possible in the first place.

The honest summary: AI helped produce a real, carefully checked piece of new mathematics, remarkably fast. It solved one accepted version of a famous question, not the harder version most people had in mind, and the actual $1 million cheque is nowhere close to being written. That's not a letdown — it's just what 'genuinely impressive but not yet finished' looks like, once you take the caveats out of the small print and put them where you can see them.