Anthropic's Levent Alpöge says he found a Riemann hypothesis proof with Claude, published only in Lean
Terence Tao calls it an enormous Lean-only proof, and the same morning Ravi Vakil rushed out a rival proof built with no AI at all.
On November 20, 2026, Levent Alpöge, a mathematician who lists Anthropic on his own homepage, posted a four-page note on the same site. In it he says he believes that, working with Claude, he has found a proof of the Riemann hypothesis, the most famous open problem in mathematics.
He says the method goes further, to the Generalized Riemann Hypothesis, and that he expects to reach the Extended version with more time. He hasn't posted a paper. The proof exists only as Lean code, split into 14 archive files on his site.
- Who
- Levent Alpöge, with Claude
- Claim
- a proof of the Generalized Riemann Hypothesis
- Form
- a Lean-only proof
- Clay prize
- at least two years after publication
What Alpöge says happened
Lean checks each step of a proof mechanically. Alpöge writes that turning his into a huge human-readable document would be routine, but that it would be unreadable, so he chose to post only the Lean. He calls the argument incredibly long and expects it to get much shorter, probably fast.
His account of the work is generous to Claude. He says he'd planned for this problem since his student days and steered the two of them toward the right setting, and then Claude overtook him near the end. A timeline he says Claude put together runs to 20 steps, from October 6, 2026 to the last one, GRH, on November 16.
He also takes up the timing, a day after Anthropic's shares started trading. The pieces came together late on a Sunday night, he writes, and he messaged Dario to ask whether he could announce. He calls the timing horrible, says there's an obvious cynical reading of it, and says that reading is false.
What mathematicians said first
Terence Tao covered it on his blog the same day, in a post about three AI-assisted results that he ranked by how far each had been digested. Two are Alpöge's joint work on blowup for the Euler and Navier-Stokes equations, and those come with human-readable papers. GRH came last.
Levent Alpöge, using an internal AI model, has generated an (enormous) Lean-only proof of the Generalized Riemann Hypothesis.
Tao sent readers to a thread on the Lean community's Zulip chat to follow the checking. Answering a question in the comments, he said the usual view is that analytic methods alone can't settle the hypothesis, and that the promising routes run through automorphic forms and arithmetic geometry. His early read is that the proof leans on all of them.
Not every reply was warm. Juliette Bruce quoted back at Tao his own September warning about turning a famous problem into a viral social media post.
It feels like Alpöge’s actions around GRH have seen your fears come to fruition.
A human-only proof turned up the same day
Ravi Vakil posted what he called a hurried announcement on his own blog that morning. Alessandro Masullo has a proposed proof of GRH, he wrote, and Brian Conrad has spent months checking it. The plan had been a year at Stanford to polish it before going public. Vakil decided to put the current draft out now.
It is all human, with no AI.
So by lunchtime there were two claimed proofs of the same conjecture, built in opposite ways. Tao added that Masullo's draft looks closer to Deninger's strategy than to Alain Connes's, and that it will take time to digest as well. One commenter pointed out that it still has placeholder footnotes in it, such as "Insert statement".
What would change outside mathematics
There's money on it. Anthropic's own August post on an earlier Claude result describes the Riemann hypothesis as carrying a million-dollar bounty, the Clay Mathematics Institute's prize. Clay's rules won't let either claim collect soon. The institute considers a solution only two years after it appears in what it calls a Qualifying Outlet, and only once it has general acceptance in the global mathematics community. A Lean archive on a personal website hasn't started that clock.
The bigger change is in who does the checking. A Lean file can be built by anyone with a large enough computer, which is a different kind of review from a referee reading a paper for months. In the comments under Tao's post, Manvir Jaswal wrote that people with powerful servers should finish building it soon, with a closer check over the following days. What a build can't settle is whether the final Lean statement really says GRH, and that part still needs number theorists.
Three months ago Anthropic was much more modest. That August post reported Claude raising a lower bound on the share of zeta zeros on the critical line from 41.6% to 67.2%, with Alpöge one of two staff mathematicians who examined the work.
We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis.
I read Alpöge's note as a claim made in good faith that nobody outside his circle has verified yet. He says he intends to write the argument up properly and to speak about it. He didn't give a date.
More on Anthropic
- Anthropic prices its IPO at $83 a share, above its range, as Nasdaq trading startsNovember 19, 2026
- Steven Pinker and Beff Jezos sign CAIS's statement that Anthropic is not a safe AI companyNovember 18, 2026
- Anthropic cuts its bug disclosure deadline to 30 days, saying another AI will find the bugNovember 16, 2026
- Anthropic says Mythos got Claude running on AMD chips in three days, almost unaidedNovember 12, 2026