Levent Alpöge says he has, “working with Claude, discovered a proof of the Riemann Hypothesis,” and that he and the model have “formalized the arguments without assumptions.” The claim is set out in a short note dated 20 November 2026 on his website, not in a journal paper, and it has not been peer-reviewed. Scientific American, The Verge and Nature describe Alpöge as a mathematician at Anthropic. The Verge and Nature report that Anthropic has not commented.
What has moved since the note appeared is a set of machine checks, not a human reading of the argument. Andrew Sutherland of MIT has published a verification page, last updated at 22:05 UTC on 20 November, of checks “run with Claude Code.” As of that update, most of the listed tests had passed, including compilation of all 175,187 modules in Lean v4.33.1 with Mathlib v4.33.1. Two headline runs — the lake comparator end to end, and the Lean-kernel replay of the exported proof — were still in progress.
What the MIT checks of the Riemann hypothesis formalisation have finished
Sutherland’s summary is that the published Lean source compiles without errors; that the five headline theorems depend only on the three standard axioms, propext, Classical.choice and Quot.sound; that every module has been replayed through the Lean kernel; that two independently written kernels, nanoda and con-leche, accept the full exported proof; and that lean4lean accepts every compiled module. Five of those lean4lean modules, he notes, hit a default step limit and passed when the limit was raised. con-leche’s --compare mode, an independent implementation of Comparator’s statement matching, “confirms all five theorems against the challenge.”
The page is also explicit about what is still running. The Lean default and “paranoid” kernels on the export had no verdict yet. The lake comparator run had already passed a statement comparison and an axiom check in one run, with the kernels still in progress. Alpöge, in the note, calls the argument “incredibly long at the moment” and says that while the formalisation is finished, he intends “to understand and impart understanding of this argument… to write it.” No such paper has been posted.
The theorems being checked include riemannHypothesis in the submission’s Main.lean, using Mathlib’s own definition of the Riemann hypothesis, “and four further theorems, including the analogous statement for all Dirichlet L-functions.” Sutherland warns that “GRH” here “means Dirichlet L-functions only.” That is narrower than some uses of the phrase “generalized Riemann hypothesis,” and it should not be stretched.
The page also carries caveats that belong next to the pass marks. The build ran on cloud machines “without a sandbox,” compiling each file by calling lean directly rather than through lake build. A 77-module sample of the output was byte-identical to a real lake build; the full set has not been compared against one. “A passing check shows that the formal statements, as written in Lean against Mathlib’s definitions, follow from the three standard axioms according to the kernels used. It relies on those kernels and on Mathlib’s definitions being what they are intended to be.”
Scientific American, in a 20 November report by Clara Moskowitz, quotes Sutherland as saying: “Honestly the chance that anything is fishy here is approximately zero.” The same piece quotes Kevin Buzzard of Imperial College London as still awaiting “a successful comparator run.” Those are their remarks, offered while two of Sutherland’s listed runs were unfinished.
The episode sits close to earlier TSN coverage of machine-checked mathematics, including OpenAI’s batch of AI maths results, some of them Lean-checked and Axiom’s account of agents formalising the simple-groups proof in Lean. Claude is Anthropic’s model; TSN’s most recent company-side account is of Anthropic’s IPO roadshow.
Most Lean checks are not a published proof
It would be easy, and wrong, to treat a stack of green marks on a verification page as the Riemann hypothesis being settled. Alpöge’s own wording is “I believe I have… discovered a proof,” posted as a short personal note. There is no journal paper, no refereeing, and, by his account, not yet a human-readable write-up of an argument he himself calls incredibly long. Machine checks can show that formal statements of the Riemann hypothesis, as written, follow from named axioms according to named kernels. They cannot, on Sutherland’s own caveat, do more than that: they rely on those kernels, and on Mathlib’s definitions being what they are intended to be.
The unfinished work is not a footnote. The MIT page still has the lake comparator and the Lean-kernel replay of the export in progress, and Buzzard has said he is waiting for a successful comparator run. con-leche’s compare mode has already accepted the five theorems against the challenge, which is a substantial independent check, not a substitute for the runs that remain. The narrower meaning of “GRH” on the page is another place where a headline can run ahead of the file: Dirichlet L-functions only, not every common use of that name.
None of this requires a reader to dismiss Alpöge’s Riemann hypothesis announcement. It should stay unverified until there is a paper the community can read, and until the remaining checks finish. Anthropic, for its part, has not added a comment of its own.
Sources
- Levent Alpöge, “remark” note on the Riemann hypothesis claim, 20 November 2026
- Andrew Sutherland, MIT, “Verification files for the Lean proof of the Riemann Hypothesis,” last updated 20 November 2026, 22:05 UTC
- Clara Moskowitz, “The AI Riemann hypothesis proof is almost certainly true, independent mathematicians attest,” Scientific American, 20 November 2026
- Robert Hart, “Anthropic mathematician claims to have solved prestigious Millennium Prize problem with Claude,” The Verge, 20 November 2026
- Davide Castelvecchi, “Mathematician says he and AI solved Riemann hypothesis — the field’s biggest outstanding challenge,” Nature, 20 November 2026
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- Axiom Says Its AI Agents Formalised the Simple Groups Proof in Lean is the closest recent TSN account of a Lean-formalised mathematics claim, and is the next read.
- OpenAI Publishes 372 AI Maths Results; About 42% Are Lean-Checked
- Anthropic Starts IPO Roadshow at an Expected $70–$78 a Share
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X

