Less than 48 hours after an Anthropic mathematician said he and Claude had proved the Riemann Hypothesis, independent mathematicians have checked the work — and certified it correct. The Lean and Mathlib communities, which maintain the proof-checking language the result is written in, issued a joint statement on 21 November 2026: “We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement. We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct.”
That turns yesterday’s story on its head. When Levent Alpöge posted his claim on Friday 20 November, the honest headline was that nothing had been verified yet. This article is the follow-up: what the independent checks actually confirmed, what each check rules out, and what is still not settled — including the $1 million Clay prize.
What changed overnight
Alpöge’s claim arrived as a four-page note on his personal website, alpo.ge, with a 14-part archive of Lean code: 175,187 files, roughly 17.4 million lines, built over 41 days of work with Claude between 6 October and 16 November. There was no conventional paper, and Alpöge himself said he does not yet fully understand every argument Claude used.
The verification was run in public by Andrew Sutherland, a mathematician at MIT, with help from Claude Code, and tracked on a live page that was still being updated at 03:10 UTC on Sunday 22 November. The checks ran in three waves: first the code compiled without errors; then independent proof-checking programs replayed it; then a separate arithmetic cross-check confirmed the result from an entirely different direction. The slowest single check — a full end-to-end run of Lean’s own comparator on the 62 GB main export — is still in progress, but every check that has finished has passed.
What the independent checks actually confirmed
The table below is drawn from Sutherland’s verification page as of 03:10 UTC on 22 November 2026. Each row is a different way the proof could have failed — and what it would have meant if it had.
| Check | Result | What it rules out |
|---|---|---|
| All 175,187 Lean modules compile (Lean v4.33.1, Mathlib v4.33.1) | Pass, zero errors | Broken logic anywhere in the 17.4 million lines |
| Only the three standard axioms used (propext, Classical.choice, Quot.sound) | Pass | Hidden extra assumptions smuggled into the proof |
| Mathlib files byte-identical to the official cache | Pass | A doctored copy of the standard library |
| Independent sandboxed rebuild reproduces the export byte for byte | Pass | A build that only works on the author’s machine |
| leanchecker kernel replay of every module | Pass, 175,187 of 175,187 | Errors the compiler missed |
| nanoda, an independently written Rust kernel, on the full export | Accepts (27 min 42 s) | A bug in Lean’s own kernel |
| con-leche and con-ron, two more independent kernels, in “compare” mode | Accept, all five theorems | The proof proving a different statement than the one claimed |
| lean4lean, a fourth independent checker, on every module | Pass, 175,187 of 175,187 | Yet another single-tool blind spot |
| Robin1984 arithmetic cross-check (Robin and Lagarias inequalities) | Pass; unmodified Comparator says “Your solution is okay!” | A proof that is formally valid but mathematically empty |
| Full lake comparator end-to-end run on the main export | Still running (as of 03:10 UTC, 22 Nov) | — |
Two rows deserve a closer look. The “compare” checks answer the sharpest question anyone asked on Friday: does the code actually prove the Riemann Hypothesis, or does it prove something subtly different? Mathematician Kevin Buzzard of Imperial College London put it plainly in a blog post that day: “The first question which one has to ask when judging a PDF which comes with some Lean code is: ‘what is the statement of the theorem in the Lean code, and does this correspond to the actual theorem statement.’” The compare runs match the proof’s statements against the challenge statements — the ones imported only from Mathlib — and all five theorems passed, including the Generalised Riemann Hypothesis for Dirichlet L-functions.
The Robin1984 cross-check is the most human-readable piece of evidence so far. It takes the proof, combines it with an independent formalisation of two classical criteria — Robin’s and Lagarias’s inequalities about the divisor sum, which are equivalent to the Riemann Hypothesis — and derives two concrete theorems about whole numbers, with no mention of the zeta function at all. Those theorems check out too, and the unmodified Comparator tool accepted the combined result at 03:31 UTC on 22 November.
“The Lean proof has been compiled and independently rechecked by the Lean kernel with only the three standard axioms. Honestly the chance that anything is fishy here is approximately zero.” — Andrew Sutherland, MIT, to Scientific American
What “verified” does not mean yet
Certified correct is not the same as finished. Four things are still open, and each matters.
No human has read the proof. The verification is machine checking: computers confirming that every logical step follows. No mathematician has yet read and understood the argument as a whole, and Alpöge has said he plans to write the human-readable version himself. Terry Tao, who blogged a preliminary analysis on Friday, noted that understanding will take time. Until that paper exists and survives peer review, the proof is correct but not yet understood.
The $1 million Clay prize is not won. The Clay Mathematics Institute still lists the Riemann Hypothesis as unsolved, as of 22 November. Its rules require a solution to be published in a qualifying peer-reviewed journal, to remain available for at least two years after publication, and to win general acceptance across the mathematics community before the money is considered. A personal website post and a code archive do not qualify, however thoroughly the code has been checked.
The scope is narrower than “all of mathematics.” The Generalised Riemann Hypothesis proved here covers Dirichlet L-functions — every modulus and every character. Dedekind zeta functions, Hecke L-functions and automorphic L-functions in general are not among the statements that were checked. That is still an enormous result; it is not the whole landscape.
One check is still running. The full end-to-end lake comparator run on the main 62 GB export, and the Lean default and paranoid kernel replays on the export, were still in progress at 03:10 UTC on 22 November. They are single-threaded and slow; the same comparator already passed on the Robin1984 cross-check, and the statement-matching portion passed in a separate run. Nothing in the remaining runs is expected to change the outcome, but they have not finished.
Why this matters beyond mathematics
This is the first of the seven Millennium Prize Problems to receive a machine-checked proof — and the proof was found with an AI system as the working partner. The work ran for 41 days, from 6 October to 16 November, with Claude assembling the argument step by step; the key steps are dated 15–16 November, and Alpöge consulted Anthropic chief executive Dario Amodei before announcing.
The honest framing matters here. Claude did not “solve” the problem alone, and Alpöge did not solve it alone either. A human mathematician steered the search, chose the strategy — adapting Stepanov’s method from function fields to integers, with theta lifts supplying a missing step — and Claude produced the formal argument at a scale no human could write by hand: 175,187 files, 250,000-plus pages of prose equivalent. The verification, too, was run with Claude Code’s help. The story is a human–AI collaboration that neither side could have done alone, checked by machines that neither side controls.
There is also a rival claim to watch. On the same day, Stanford’s Ravi Vakil released an all-human draft by Alessandro Masullo, also claiming the Generalised Riemann Hypothesis, which had been under private review for months. That draft has not been through the same public verification, and the two claims are independent — if both hold, the mathematics community will have two proofs, not one. Neither has been peer-reviewed yet.
What happens next
The sequence from here is slow by design. First, Alpöge writes the human-readable paper — the version mathematicians can actually read, argue with and teach. Then peer review, which for a result of this size will take months at minimum. Then, if the community accepts it, the Clay Institute’s two-year clock starts. The $1 million, if it ever comes, is years away.
Meanwhile the verification page keeps updating, the remaining comparator run will finish, and mathematicians will start mining the proof for the ideas inside it — the new techniques, not just the headline. The proof’s real legacy may be the method: a working demonstration that an AI system can help build a formal argument at a scale that settles a 167-year-old problem, and that independent machines can then check it to a standard no human referee could match.
Yesterday we wrote that the claim was extraordinary and unverified. Today the verification is in, the certification is public, and the honest headline has changed: the proof is correct. What is not yet settled is whether anyone understands it — and whether the mathematics community will accept it as a proof in the human sense of the word.
Frequently Asked Questions
Has the Riemann Hypothesis been proved?
Yes, in the formal sense: the Lean code compiles, three independent proof-checking programs accept it, and the Lean and Mathlib communities have certified both the statement and the proof as correct. What is still missing is a human-readable paper and peer review — no mathematician has yet read and understood the full argument.
Has Alpöge won the $1 million Clay prize?
No. The Clay Mathematics Institute still lists the Riemann Hypothesis as unsolved. Its rules require publication in a qualifying peer-reviewed journal, two years after publication, and general acceptance by the mathematics community. None of those steps has happened yet.
What did Claude actually do?
Claude worked with Alpöge for 41 days, from 6 October to 16 November 2026, building the formal Lean argument step by step — including the theta-lift idea that supplied a missing step. Alpöge steered the strategy and chose the method; Claude produced the 17.4-million-line formal proof. Alpöge has said he does not yet fully understand every argument Claude used.
What is the Generalised Riemann Hypothesis, and was it proved too?
The Generalised Riemann Hypothesis extends the same claim from the Riemann zeta function to Dirichlet L-functions. Yes — the verification covers five theorems, including the Generalised Riemann Hypothesis for all Dirichlet L-functions. It does not cover Dedekind zeta, Hecke or general automorphic L-functions.
What about the second, all-human proof announced the same day?
Stanford’s Ravi Vakil released a draft by Alessandro Masullo claiming the Generalised Riemann Hypothesis by an entirely human route, under private review for months. It has not been through the same public machine verification, and the two claims are independent. If both hold, mathematics will have two proofs.
Sources and Methodology
Sources: Scientific American (updated 21 November 2026, reporting the Lean and Mathlib joint statement and Sutherland’s comments); Andrew Sutherland’s verification page, MIT (last updated 22 November 2026, 03:10 UTC — the check-by-check results quoted in the table); ExplainX claimed-vs-verified tracker (updated 21 November 2026, 14:00 UTC); Clay Mathematics Institute, Riemann Hypothesis (prize rules and current listing). Background on the original claim and the Masullo draft is in our 21 November explainer; the IPO timing is covered in our 21 November sell-off report.
Method: every check in the table was read from Sutherland’s verification page on 22 November 2026 and cross-checked against Scientific American’s updated report and the ExplainX tracker; where the sources differ on timing (Scientific American says the comparator run “has now been completed”, while the MIT page, updated later, still lists the end-to-end run as in progress), this article follows the MIT page as the most recent primary record and says so. Nothing here is hands-on tested: the author did not run the Lean checks. Verification date: 22 November 2026.
Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


