ThinkFacility Sign in
  1. Home
  2. News
  3. Anthropic

Michael Rubinstein says all four Lean checks of Alpöge's Riemann proof with Claude found no errors

Four machine checks of 17 million lines came back clean, but number theorists still have to judge whether the Lean statement really is GRH.

On November 21, 2026, Michael Rubinstein posted the results of an independent check of Levent Alpöge's Lean proof of the Riemann hypothesis. Alpöge had announced the proof, found with Anthropic's Claude, the day before. Rubinstein's checks found no errors.

The page, on the University of Waterloo's math site, lists four machine checks, and every one passed. It also links to checks by three other people and to a dashboard that keeps a tally of who has verified what.

Checked by
Michael Rubinstein, with Claude Code
Proof size
175,187 files, about 17 million lines
Result
All four checks finished with no errors
Build time
3 h 18 min

What the four checks cover

Lean only accepts a proof whose every step goes through its kernel, the small core program that does the checking. Rubinstein compiled all 175,187 modules and then listed the axioms the five main theorems rest on. They need only Lean's three standard axioms, so nothing in the code quietly assumes what it's meant to prove.

Two kernel checks came next. A replay of every module through Lean's own kernel accepted all of them. A second kernel, written separately and called nanoda, checked the 1,484,729 declarations the theorems depend on. (For a bad step to slip through, both kernels would have to share the same bug.)

All four checks finished with no errors. All 175,187 modules compile.

From Alpöge/Claude proof of the Generalized Riemann Hypothesis

The build ran on university servers, mostly AMD EPYC and Intel Xeon machines, and took 3 hours 18 minutes. Rubinstein put the scripts in the repository and kept a dated log of the problems met along the way, so anyone with the hardware can rerun it.

The others are Andrew Sutherland at MIT, Elliot Arledge, who posted a rebuild and an axiom audit on GitHub, and a full kernel replay that Bartosz Naskręcki's dashboard credits to Simon DeDeo.

What a clean build doesn't settle

Rubinstein spells out what the result rests on.

this assumes only that the Lean kernel (checked twice, by independent implementations) is sound and that Mathlib's definitions of ζ and of Dirichlet L-functions mean what they are supposed to.

From GRH Lean Check

That's the gap our first brief on the claim pointed to: whether the final Lean statement really says GRH. Rubinstein's page sets out the five theorems exactly as Lean states them and explains why they say what RH and GRH for Dirichlet L-functions say. It's a careful reading. Number theorists will still go over it.

A clean build also doesn't make the argument readable. Rubinstein has posted a 164-page working draft, written with Claude, Codex and Gemini, that explains the proof for the zeta function in the style of classical analytic number theory. It gets RH from one inequality between weighted sums over primes, called (M), plus two estimates for Dirichlet characters. The draft doesn't prove (M). That part still comes from the Lean.

What Pichai and Musk said

Google's CEO posted on the afternoon of November 20, before any of the checks were in.

Sundar Pichai@sundarpichai · Google

Somewhere Euler, Gauss and Riemann in awe today!

View the post on X

Musk posted at 2:46 am Eastern on November 21, and hedged.

Elon Musk@elonmusk · xAI

Even if Riemann turns out not to be fully solved yet, it will be soon enough

View the post on X

By that afternoon Musk had gone further. A proof, the post said, removes the conditional nature of other theorems that depend on the hypothesis, and it called the result "yet another mind-blowing thing achieved by super intelligence".

What changes outside mathematics

Musk's point about conditional theorems has a plain example in computing. The Miller test tells you whether a number is prime, and if the generalized Riemann hypothesis holds it's deterministic: you only need to try bases up to about 2(ln n)². Wikipedia's article on the test is blunt about what that's worth day to day. "The Miller test is not used in practice," it says, and it's slower than commonly used proof methods such as APR-CL and ECPP. So software that tests primes won't change because of this.

Prize money won't move quickly either. The Clay Mathematics Institute, which offers a million dollars for the Riemann hypothesis, says it doesn't take submissions directly. It only considers a solution after publication in an approved outlet, a two-year wait and acceptance in the mathematics community worldwide.

What has changed is who can check a famous proof, and how fast. A claim that went up as a 14-part archive on a personal site on November 20 had been compiled and replayed by four groups within a day, and I think that's the most concrete thing to come out of it so far.

Rubinstein's explanation of the proof is dated 21 November 2026 and marked a work in progress.

More on Anthropic

All Anthropic stories