Ten days after claiming a solution to Navier-Stokes, OpenAI says it’s closing in on a second Millennium Prize problem. The Information reports that the company is near a result on another of the seven famous math challenges, each carrying a $1 million bounty from the Clay Mathematics Institute. OpenAI hasn’t named the problem publicly, and it hasn’t said when it plans to show its work.
That’s the whole story on the surface. The context underneath is where it gets interesting.
What we know so far
Here’s the sequence, pieced together from The Information, OpenAI’s own statements, and coverage from CNN, Nature and others:
- On September 8, OpenAI announced that its system had produced a proof for the Navier-Stokes existence and smoothness problem, a 200-year-old question about whether fluid-motion equations always behave. The run took 88 hours and used as many as 10,000 AI agents working in parallel.
- OpenAI said it won’t claim the Clay prize for that result. The Clay Institute hasn’t verified the proof, and the broader math community is still working through it.
- Shortly after, OpenAI told The New York Times it had made “substantial progress” on a second Millennium problem. No name, no timeline.
- Now The Information reports the company is close to actually solving that second one.
Online speculation points at the Hodge Conjecture, a question in algebraic geometry about which shapes can be built from simpler algebraic pieces. OpenAI hasn’t confirmed that, so treat it as a rumor until they say otherwise.
Why the first claim is still contested
The Navier-Stokes announcement came with baggage. According to reporting around the release, OpenAI pointed its system at the problem on September 1 after hearing that two mathematicians, Levent Alpöge and Tristan Buckmaster, had cracked a version of it using Anthropic’s models. That kicked off a priority dispute over who got there first and how much of the work was genuinely new.
There’s also the verification gap. A proof from 10,000 agents running for 88 hours is not something a referee reads over a weekend. Mathematicians need to trace every step, and a single unpatched hole sinks the whole thing. The Clay Institute has a two-year waiting period after publication before it even considers awarding a prize, so “solved” and “accepted” are very different words here.
So when OpenAI says it’s close on problem number two, the honest read is: close to producing a candidate proof, not close to a settled result.
Why this matters beyond math
What stands out to me is the shift in what these labs are using their compute for. A year ago the headline was gold-medal scores on the International Math Olympiad. Those are hard problems, but they’re problems with known answers that a human can check in an afternoon. Millennium problems are open research questions. Nobody knows the answer. If the proofs hold up, that’s a different category of capability.
A few implications for practitioners:
- Agent swarms are the method. Ten thousand agents on one problem for days is the template. Expect labs to push this pattern into drug discovery, materials science and chip design, where the search space is huge and verification is expensive.
- Verification becomes the bottleneck. Generating candidate proofs is now cheap. Checking them isn’t. Tools like Lean and other formal proof assistants are about to get a lot more attention and funding.
- The race is now a two-lab race. The Alpöge and Buckmaster story shows Anthropic’s models were in the mix on Navier-Stokes. Whatever OpenAI is chasing next, assume competitors are chasing it too.
- Marketing is running ahead of peer review. OpenAI has now made two big math claims in ten days, neither independently confirmed. That’s a pace the math community isn’t built to match, and the friction is showing.
What to watch
The real test is whether OpenAI publishes the second proof in a form mathematicians can actually check, and whether it survives that scrutiny. Until then, this is a promising signal and a very loud one, not a verified breakthrough.
I’d also watch how the Clay Institute responds. It hasn’t said much about the Navier-Stokes claim yet. Two AI-generated candidate proofs landing in the same month puts pressure on an institution built around a slower era of mathematics.
The Information has more on the internal push behind the second problem.