Compiled from information available as of 1:00 p.m. on November 21, 2026 (Japan Standard Time).
Levent Alpöge, a mathematician at Anthropic, published a document on his personal website on November 20, 2026 (US time) saying that, working with the AI model Claude, he believes he has found a proof of the Riemann Hypothesis. The proof was released as code for the Lean proof assistant. Andrew Sutherland of MIT and others have independently machine-checked that code and report that it passes the main checks. There is no human-readable paper yet, and mathematicians have not yet reviewed the mathematics itself.
Alpöge announces a "proof" made with Claude, releasing only Lean code
In a four-page document published on November 20, 2026 (US time), Alpöge wrote: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He says the method extends to the Generalized Riemann Hypothesis (GRH), and that near the end Claude "far outpaced" him. The proof consists of 175,187 Lean 4 files (about 1.4 GB), distributed as a 14-part archive. He explains that the argument is currently very long, and that since posting a huge, unreadable document would be inappropriate, he decided "to only present Lean." The announcement was made on Alpöge's personal site. According to Nature, Anthropic did not immediately respond to its request for comment.
Sources: alpo.ge (Alpöge's document, Nov 20) nature.com (Nov 20)
MIT's Sutherland reports passes on several independent checks; some checks still running
According to Sutherland's verification page (updated 7:05 a.m. on November 21, 2026, JST), all 175,187 modules compile without errors. The five headline theorems depend only on Lean's three standard axioms (propext, Classical.choice, Quot.sound). Every module was replayed through Lean's own kernel with zero rejections, and two independently written kernels, nanoda and con-leche, accepted the full exported proof. A separate tool also confirmed that the theorem statements match a previously published "challenge" file. At that time, the end-to-end run of the official comparison tool "lake comparator" and re-checks with Lean's default and "paranoid" kernels were listed as still running. Sutherland also notes as a caveat that the build ran on cloud machines without a sandbox. In a Scientific American article dated November 20, he said, "Honestly the chance that anything is fishy here is approximately zero."
Sources: math.mit.edu (Sutherland's verification page) scientificamerican.com (Nov 20)
Another third-party re-check on GitHub: zero failures across all modules
GitHub user luosiallen has also published a verification log, "VERIFICATION_LOG.md," covering work through November 21, 2026 (JST). According to the log, all 175,187 modules were rebuilt from scratch on a 288-core machine with zero failures. No "sorry," custom axiom declarations, or settings that skip kernel checking were found in the sources. The five main theorems depend only on the three standard axioms, and the type of riemannHypothesis was confirmed to be exactly the RiemannHypothesis definition in unmodified Mathlib, Lean's shared mathematics library. Finally, 183,877 modules were re-checked with the same replay routine used by leanchecker, finishing with zero failures (around 7:49 a.m. on November 21, 2026, JST). The author writes that this is strong evidence but not a substitute for expert mathematical review. In Japan, uehaj shared the log on X at 8:36 a.m. on November 21 (JST).
Sources: github.com (luosiallen's verification log)
What has been checked, and what has not
What has been checked: the Lean code compiles without errors, the theorem statement matches Mathlib's definition of the Riemann Hypothesis, no axioms beyond the standard ones are used, and several independent kernels accept the proof. What has not: the full lake comparator run, and humans understanding and reconstructing the argument. On November 21 (JST), Terence Tao described it on his blog as an "(enormous) Lean-only proof of the Generalized Riemann Hypothesis." He said that very initial analysis suggests it uses the full power of modern developments in the Langlands program and arithmetic geometry, and that digesting it will take some time. Sutherland notes that "GRH" here means Dirichlet L-functions only. According to a Nature article dated November 20, Andrew Granville pointed out that the announcement came just days after Anthropic's IPO and called the released document "garbled." In his document, Alpöge wrote that a cynical view of the timing is false.
Sources: terrytao.wordpress.com (Nov 21 JST) nature.com (Nov 20)
What is the Riemann Hypothesis?
The Riemann Hypothesis was proposed by the German mathematician Bernhard Riemann in 1859 and is closely tied to how prime numbers are distributed. It states that all the non-trivial points where a function called the "zeta function" equals zero lie on a single line where the real part is 1/2. If true, it would mean the primes are spread out in close to the most regular way possible, and many results in number theory have been built assuming it. In 2000, the Clay Mathematics Institute named it one of its seven Millennium Prize Problems, with a US$1 million prize for a solution. Unsolved for more than 160 years, it is widely regarded as one of the most important open problems in mathematics.
What is Lean? A system where a machine checks a proof step by step
Lean lets mathematicians write a proof as a program, and a small checker called the "kernel" verifies mechanically that each step is correct. If the kernel accepts it, the stated theorem follows from the axioms. Whether the theorem statement really says what it is meant to say, and whether the kernel itself has any holes, must be checked separately. The current checks, which compare definitions and replay the proof with several kernels, are aimed at exactly those points.
AIZNOH's view
AI is now involved both in producing the proof and in checking it, while humans have not yet read its contents: the usual order has been reversed. Sutherland's page also notes that his checks were run with Claude Code. Passing machine verification is strong evidence, but for the mathematical community to accept the result, people will need to understand and rewrite the argument. The lake comparator result and expert reading of the proof are what to watch next.