Has AI solved the Riemann hypothesis? What we know about the proof made with Claude
A mathematician at Anthropic says he proved one of math's great unsolved problems with help from Claude. On the same day, a Stanford researcher announced another proof made without AI. Here's what it is, what's been checked and what's still missing.

Something happened this week that may end up in the history books… or may not. Levent Alpöge, a mathematician who works at Anthropic (the company behind Claude), says he has proved the Riemann hypothesis, one of the most famous and hardest problems in mathematics. He did it working with Claude, his company’s AI.
To make it even stranger, on the very same day, November 20, another proof of a version of the problem was announced, this one by a human and without AI.
Here’s the story without the equations, based on TechCrunch’s reporting.
At a glance
| The problem | The Riemann hypothesis, posed in 1859 (167 years ago) |
| Why it matters | It’s one of seven “Millennium Problems”: solving it can earn a $1 million prize |
| Proof with AI | Levent Alpöge (Anthropic), working with Claude |
| Proof without AI | Alessandro Masullo (Stanford), for the generalized Riemann hypothesis; 12 years of work, he says |
| Has it been checked? | Alpöge’s proof is written in Lean, a language a computer can verify, and several mathematicians say it compiles. But experts still need to understand the argument |
| Is it official? | Not yet. No journal publication and no statement from the Clay Institute |
The Riemann hypothesis, without equations
Prime numbers (2, 3, 5, 7, 11…) can only be divided by 1 and themselves. They seem to pop up at random, but in 1859 German mathematician Bernhard Riemann suspected they hide a very precise pattern, tied to a special function (the “zeta function”).
Very roughly, the hypothesis says that all the important points of that function sit on the same line. If it’s true, we’d know very precisely how primes are spread out. Thousands of math results already assume it’s true, but nobody had proved it.
Why should you care? Prime numbers are behind the encryption that protects your card payments and messages. Don’t worry: proving the hypothesis doesn’t break that security. It gives certainty about something people already assumed.
How the Claude proof came about
According to TechCrunch, Alpöge wrote an essay describing how he found the pieces of the argument through a lot of confusion, talking things through with Claude. Which Claude model he used hasn’t been disclosed.
What stands out is how he presented it:
- The proof is written in Lean, a “proof language” where a program checks every single step. If the program accepts the whole proof, a hidden logical error is very unlikely.
- Several mathematicians on the Lean forum, including MIT’s Andrew Sutherland, say they compiled it successfully. TechCrunch later added that independent sources have checked it.
- What doesn’t exist yet is an explanation a human can read and follow. Alpöge admits he feels a bit guilty about releasing it this way and has promised to write one.
Alpöge also released, with NYU mathematician Tristan Buckmaster, another proof related to the Navier-Stokes equations, another Millennium Problem. Even less is known about that one so far.
The other proof: 12 years, no AI
The same day, Ravi Vakil, a Stanford professor and president of the American Mathematical Society, posted a proof by Alessandro Masullo, a Stanford postdoc, on his blog. It covers the generalized Riemann hypothesis, a broader version of the problem.
Vakil stresses that it’s “all human, with no AI.” Masullo says it took him 12 years. His proof isn’t written in Lean, so TechCrunch says it’s too early to know whether it holds up: experts will have to check it by hand, which can take months.
So is it solved or not?
The honest answer: it’s very promising, but not final yet. Three things to keep in mind:
- A computer accepting the proof isn’t everything. Lean checks that each step is valid, but experts must confirm that what was proved is exactly the Riemann hypothesis and not a look-alike statement.
- The prize isn’t automatic. The Clay Institute’s rules require the solution to be published and accepted by the math community for a while before the $1 million is awarded.
- There’s a possible conflict of interest. Alpöge works at Anthropic, which just went public and benefits a lot from headlines like this. That doesn’t make the proof wrong (math gets checked, not believed), but it’s one more reason to wait for review.
Why this matters for AI
Until recently, chatbots’ big flaw was that they make things up with total confidence (we explain why in AI hallucinations). Pairing AI with tools like Lean, which can’t be fooled, is exactly what lets you trust its results in math.
If it’s confirmed, it would be one of the biggest moments in AI history, right up there with the ones in the history of AI in 12 key moments. It also raises a fun question: who gets the credit, the mathematician or the machine? For now, Alpöge signs the work and describes Claude as his thinking partner.
Note: this story is based mainly on TechCrunch. We haven’t read Alpöge’s full essay, and things may change in the coming days as mathematicians keep reviewing.
Sources
- TechCrunch: AI claimed a solution to the 167-year-old Riemann hypothesis, and a human did, too (Nov 20, 2026)
- Levent Alpöge's essay on his proof (cited by TechCrunch)
- Ravi Vakil: announcement of Alessandro Masullo's proof (Nov 20, 2026)
- Clay Mathematics Institute: the Riemann hypothesis, a Millennium Problem

