Back to Explore
NewsVibeFix editorialUpdated Nov 21, 2026

Claude and a Mathematician Claim the Riemann Hypothesis: 250,000 Pages Unwritten, 17.4 Million Lines of Lean Machine-Checked

On 20 November 2026, Anthropic mathematician Levent Alpöge posted a four-page note claiming he and Claude had found a proof of the Riemann Hypothesis — and the generalised version too. The proof exists only as Lean code: 175,187 files, about 17.4 million lines. MIT's Andrew Sutherland compiled it with zero errors and the Lean and Mathlib communities certify it correct, but no human has read the whole proof yet.

A man seen from behind stands before a chalkboard covered in handwritten mathematical formulas, raising his right hand to write with chalk

In one sentence: the hardest problem in mathematics suddenly has an answer that only machines have read

On 20 November 2026, Levent Alpöge, a mathematician at Anthropic, posted a four-page note titled "Remark" on his personal website, and its first sentence woke up the whole mathematical world: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." In the same breath he added one more claim: the method also settles the more general Generalized Riemann Hypothesis. There was no paper and no peer review — only a four-page note and a Lean formalization of the proof split across 14 compressed archives: 175,187 files, about 17.4 million lines of code. By the evening of 21 November, the Lean and Mathlib communities had published a joint statement certifying that both the statement and the proof are correct, and Scientific American's update the same day reported that the key comparison check had finished. But the conclusion comes first: this is a claim in which machines have said yes while humans have not yet understood it, and there is still a road to travel before anyone can say the mathematical community accepts the Riemann Hypothesis as proved. This article keeps those two layers apart.

This deserves its own article not only because of the name on the problem. It is the first time anyone has claimed a Millennium Prize Problem with a proof that only machines have read, and the first time the question of what role AI plays in mathematics has moved from "helping with the arithmetic" to "helping find the idea". Alpöge writes it without disguise in his note: near the finish line, Claude "far outpaced me. No, Claude flew, and danced, and sparkled near that finish line." For people who use AI to write code every day, the shock is direct: a model did not just complete code or fix tests; it took part in a 41-day assault across the whole of number theory, and at the decisive point it proposed an idea its human author admits he had not seen coming. Whether that idea ultimately stands is something machine checking has begun to answer; human understanding is still on its way.

First, the problem itself: what the Riemann Hypothesis actually asks

The Riemann Hypothesis was proposed in 1859 by the German mathematician Bernhard Riemann, and it asks about the pattern behind the prime numbers. Primes are the numbers divisible only by 1 and themselves: 2, 3, 5, 7, 11, and so on. They grow sparser along the number line, yet show no obvious pattern. Riemann translated the question into one about the Riemann zeta function: do all of its "nontrivial zeros" lie on one particular vertical line — the line where the real part equals one half, the so-called critical line? If the answer is yes, the distribution of primes is governed by a very strong constraint, and many theorems about primes can be improved dramatically; if the answer is no, much of modern number theory has to be rebuilt from the foundations.

The conjecture has stood unsolved for 167 years, from 1859 to today. In 2000 the Clay Mathematics Institute listed it among its seven Millennium Prize Problems, with one million dollars for each solution, and none of the seven has yet been officially recognised as solved. Its generalised version goes further still: it asks not only about the Riemann zeta function but about a whole family of related Dirichlet L-functions, whether they all obey the same rule. What Alpöge claims is both at once. His note adds one more line: with more time, he expects the still broader Extended Riemann Hypothesis to fall too. That is a large claim, but this time he is holding something machines have already checked, not an empty promise.

What this "proof" actually looks like: four pages, plus 17.4 million lines of code

Start with the shape of the release, because the shape is itself news. Alpöge published no paper. He published two things: a four-page note, and a formal proof in Lean. Lean is a proof-assistant language in which mathematicians write theorems and proofs as code that a machine can check line by line; code that compiles without errors is usually taken to mean that every step of the logic holds. The note explains why there is no paper: written out as prose a human could read, the argument already runs to more than 250,000 pages. Alpöge calls that draft "slop" himself and saw no point in posting it, so he posted only the Lean code. He is candid, too, that he has not yet mastered all of the mathematics involved, and that he intends to work at understanding the argument, talking about it, and writing it up properly.

The outline of the idea in the note is, in his words, a line "which will suffice to make fellow mathematicians fall out of their chairs". The argument descends from Stepanov's method: Stepanov used a divisor carrying multiplicity to settle the Riemann Hypothesis over function fields, but over the integers multiplicity is hard to find — it corresponds to something like differentiating whole numbers. Alpöge says he steered Claude toward two ideas that appear together in a toy setting, and Claude's key step was to chain theta lifts: for a von Mangoldt sum of length X, Claude proposed a chain of theta lifts of length about the square root of X, each level mimicking one unit of multiplicity, stacking up until the missing multiplicity is recovered. The timeline in the note has 20 steps, running from the Stepanov determinant on 6 October 2026 to the iterated theta lifts settled late on 15 November and the Generalized Riemann Hypothesis on 16 November — 41 days in all. Alpöge says he set the direction while Claude pushed the technique forward, and he admits frankly that in the parts where Claude is the expert and he is not, he is paraphrasing a model-written summary that "might make zero sense", and that "the Lean should suffice to clarify everything".

The scale alone tells you this is not an ordinary release: 14 archives totalling 239.3 MB compressed, unpacking to 175,187 Lean files and about 17.4 million lines of code, with the exported proof data running to 61.6 GB. Anyone who wants to check it themselves will need, according to Sutherland's verification page, roughly 100 GB of RAM and 70 GB of disk. This is no longer a "download it and run it" exercise; it is a verification project that needs a cluster.

Timeline from 6 October to 21 November 2026: work begins on 6 October, the iterated theta lifts are settled late on 15 November, the Generalized Riemann Hypothesis follows on 16 November, the four-page note and Lean code go public on 20 November, and the Lean and Mathlib communities certify the proof as correct on 21 November

The timeline is worth pulling out on its own, because it answers a question many readers will ask: what actually happened in those 41 days? On 6 October, the Stepanov determinant; through mid-to-late October, the team moved back and forth between the identity qN = Mk + R, trace formulas and the relative trace formula; late on 15 November, the chain of iterated theta lifts finally connected, and Alpöge wrote in his note that he messaged Anthropic chief executive Dario Amodei that night to ask whether he could announce; on 16 November, the Generalized Riemann Hypothesis part fell into place; on 20 November, the four-page note and 14 archives went live; on 21 November, the Lean and Mathlib communities published their joint statement. The whole thing moved at a speed that is abnormal by the standards of mathematics — a Millennium problem going from "an idea" to "a proof" is normally measured in years, and here it took six weeks. That speed is exactly what excites supporters and unsettles sceptics: it is both evidence and grounds for doubt.

How the machines checked it: five gates, each harder than the last

The most solid part of this story is the verification carried out between 20 and 21 November, led by Andrew Sutherland of MIT with the whole Lean community pitching in. Sutherland published the result of every check on a public verification page at MIT, and they are worth going through one by one, because they answer five different questions, each of which was once a reasonable doubt.

Gate one: does the code compile? All 175,187 modules compiled under Lean 4.33.1 and Mathlib 4.33.1 with zero errors, finishing at 16:43 UTC on 20 November. Compiling is only the entry ticket, but zero errors across more than 175,000 modules already rules out the possibility that the code simply does not run.

Gate two: did the proof smuggle in extra assumptions? In Lean you can run "#print axioms" to see which axioms a theorem depends on. The five headline theorems depend only on the three standard axioms — propext, Classical.choice and Quot.sound — with no additional assumptions. The environment contains 15 axioms in total, all from Lean's core library. This rules out a hidden premise nobody noticed.

Gate three: is what was proved actually the Riemann Hypothesis, rather than a proposition that merely looks like it? This is the easiest place to cheat: you can always prove a "Riemann Hypothesis" you redefined yourself. Sutherland's page states that the five headline theorems use Mathlib's own definition — every zero of zeta other than the trivial zeros has real part one half — and that the generalised version covers all moduli and all Dirichlet characters' L-functions. The community added a file that isolates the statement while importing only Mathlib, then compared it against the theorem Alpöge proved using two independent tools, comparator and con-leche --compare, confirming that the proved theorem is exactly Mathlib's statement of the Riemann Hypothesis.

Gate four: check it again with a different kernel. Lean's own kernel replay (leanchecker) passed all 175,187 of 175,187 modules; nanoda, an independent kernel written in Rust, accepted the full export in 27 minutes 42 seconds on 80 threads; con-leche accepted 1,468,418 declarations in 3 minutes 15 seconds on 64 threads; lean4lean passed all 175,187 of 175,187 modules one by one, with five modules hitting its default step limit and passing when rerun with the limit raised. The point of using several kernels is that even if Lean's own kernel had an unknown bug, the other kernels are unlikely to share the same bug.

Gate five: an arithmetic cross-check that does not depend on the zeta function at all. Sutherland combined the proof with Robin1984, an independent formalization that states equivalents of the Riemann Hypothesis as the Robin and Lagarias inequalities — statements about divisor sums, with no mention of zeta. The two combined inequalities compiled and passed the same multi-kernel checks. The value of this gate is that it does not depend on whether Mathlib's definition of zeta matches intuition; it lands directly on elementary inequalities about divisor sums.

One gate still had a loose end as this article was written: the official lake comparator run over the full export. Sutherland's page, in its snapshot at 12:55 UTC on 21 November, said the comparator end-to-end run and the Lean kernel replay of the export were still in progress; the same page also records that the comparator had already passed for the Robin-combined export, printing "Your solution is okay!". Scientific American's update at 16:45 UTC on 21 November, however, says that the comparison run "has now been completed", and the Lean and Mathlib communities' joint statement that evening certifies directly that "both the statement and the proof of the Riemann Hypothesis are correct". The half-day gap in wording comes from checks finishing one after another during 21 November, with different pages refreshed at different times. This article follows the later joint statement and Scientific American update, and records the timing gap honestly in the doubts section at the end of the evidence file.

Why this time is different: the summer already sounded three warnings

Seen only on 20 November, this looks like a bolt from the blue; placed back on the timeline of the second half of 2026, it is the fourth thunderclap. The first came in July, when Alpöge posted a Claude-generated counterexample to the Jacobian conjecture on social media, shaking up algebraic geometry. The second came in August, when an unreleased version of Claude reportedly raised the lower bound on the proportion of zeta zeros on the critical line from the old record of 41.6% to two thirds, with Alpöge and fellow Anthropic mathematician Ralph Furman involved in the work and a Lean formalization passing the comparator check. The third came in October, when OpenAI published a proof of a weaker statement — the quasi-Riemann Hypothesis — with more than 700 results and no human authors listed. Each step pushed the line "can AI touch the hardest problems in mathematics?" one notch forward; 20 November was simply the first time it reached the hardest one of all.

One more piece of background matters: Alpöge is not someone who appeared out of nowhere as an "AI mathematician". He studied at Harvard as an undergraduate, earned his PhD at Princeton under Manjul Bhargava, was a Junior Fellow at the Harvard Society of Fellows, and joined Anthropic's technical staff after his PhD in 2020. In his note he writes that the Riemann Hypothesis has been his principal mathematical goal since childhood, and that he deliberately set out to learn every tool he might one day need for it. His role in this collaboration was not "hand the problem to the model and wait for an answer": he chose the direction, combined two ideas from a toy setting, and judged which paths were worth pursuing, while Claude did the technical work of walking the path — and near the finish line, in his words, "far outpaced" him. That division of labour — humans set strategy, models run the tactics — is exactly the point Terence Tao picked out in his commentary.

Tao's cold water: of three results, this one is the hardest to digest

Fields medallist Terence Tao followed up on his personal blog on 20 November, the same day. He ranked three AI-related mathematical results that had appeared around the same time in "decreasing order of how 'digested' the results currently are": the Navier–Stokes result came first, because it has both a readable paper and a Lean formalization; the Riemann Hypothesis claim came last, because it is a Lean-only proof with no paper. His words: "I imagine the digestion process will take some time, though in this case at least we have a human author to communicate with to assist with this task."

That sentence is worth taking apart word by word. First, "a Lean-only proof" is not an insult; it is a precise description of the state of things: the proof exists, but only in machine-readable form. Second, "the digestion process will take some time" is Tao reminding everyone that machine checking is not the same as human understanding. The mathematical community has never accepted a theorem merely because a kernel said yes; someone has to be able to explain it, to see why it is true, and to see what else it can do. Third, "at least we have a human author to communicate with" is what makes this proof luckier than a purely model-generated output: Alpöge is alive, present, and has promised to write the argument up. Many AI-generated mathematical results do not have that luck — their author is the model itself, and nobody can explain on its behalf.

Tao's ranking also throws light on a second story from the same day: Ravi Vakil, president of the American Mathematical Society, posted a draft by a human mathematician claiming the Generalized Riemann Hypothesis without AI, which is reportedly undergoing major revision. Two claims landing on the same day shows how heated the race in this direction has become — who posts first matters. This article does not go into that draft, because its primary source has not been read first-hand; it is left for later verification.

My judgement: three layers, believed separately

Layer one, formal correctness: this can be believed, and believed strongly. 175,187 modules compiling with zero errors, only the three standard axioms, four independent kernels accepting the result, the statement confirmed by two comparison tools, plus an arithmetic cross-check that never mentions zeta — that combination leaves very little room for a back door hidden in the code or a switched statement. When Sutherland told Scientific American that the chance "anything is fishy here is approximately zero", it was not politeness; it was a judgement built on the checks above. The joint statement from the Lean and Mathlib communities goes further still: "We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct."

Layer two, mathematical understanding: there is none yet, and honesty requires saying so. Alpöge himself admits he has not yet mastered all of the mathematics in the argument; Tao says digestion will take time; no human has read and understood the whole proof. What we know now is that "there is a 17.4-million-line argument that derives Mathlib's statement of the Riemann Hypothesis from Mathlib's axioms" — not "we understand why the Riemann Hypothesis is true". The gap between those two sentences is the gap between "machines can check it" and "humans can understand it", and it is the real work of the coming months, or years: compressing this proof, simplifying it, and turning it into something a person can tell. Alpöge himself predicts in his note that simplifications will follow "likely immediately".

Layer three, formal recognition: that process has not even started. The Clay Mathematics Institute's one-million-dollar prize requires a proof that has been peer-reviewed and accepted by the mathematical community, not a Lean archive. There is no rule in the prize process that says "machine checking counts as acceptance". In other words, even if every check passes, the Riemann Hypothesis remains officially "unsolved" until human mathematicians read it, write it up, and pass it through peer review. How long that takes, nobody knows; optimistic estimates say months, pessimistic ones say years. But whatever happens, 20 November 2026 has already changed one thing: from now on, "AI proved a hard mathematics problem" is no longer a hypothetical discussion. It is a reality with concrete files, concrete check records, and concrete community backing.

For people who use AI to write code every day, there is one more lesson closer to home. Sutherland's verification page carries a small line that is easy to miss: the checks were run "with Claude Code". The verification itself was a human-machine collaboration: Claude helped find the proof, and Claude Code helped run the checks. Alpöge ends his note by hoping the argument "provides a glimpse towards the miracles that lie ahead, and shows us the joy and wonder we are about to find in the simplest things". That sounds lyrical, but in a story about a 167-year-old problem falling in 41 days, it is not an exaggeration. The real question is not "will AI replace mathematicians", but "when AI can walk in six weeks a road humans did not walk in decades, what do mathematicians' working habits, the shape of a paper, and the peer-review system have to become?" This proof, read only by machines, has dragged that question out of the journals and put it in front of everyone.

How this relates to what we have already published

Between 19 and 20 November we published two stories directly connected with Anthropic: a full record of its first week on the stock market — priced at $83, opened at $111.50, fell back to $88.91 in two days — and OpenAI's anti-distillation research on the GPT-6.2 safety defences. This article shares only a coincidence of timing with the first: Alpöge's note appeared the day after Anthropic's listing, and he admits in the note itself that "the precise timing is horrible", going out of his way to say the cynical reading is false; none of the IPO figures are repeated here. The second story is a different subject entirely — model safety versus a mathematics proof.

One more boundary needs drawing: we have never before published on the Riemann Hypothesis, the Generalized Riemann Hypothesis, or Lean formalized proofs. The topic did appear in our cycle231 dropped list, but the reason it was cut then was precisely that only second-hand summaries had been seen and no primary source had been verified; this round the four-page note has been read in full, as have Sutherland's verification page, the Lean and Mathlib communities' joint statement, and Scientific American's 21 November update, so the reason for cutting it no longer stands.

A final word of advice for anyone tempted to forward this story as "the Riemann Hypothesis has been proved": add the word "claimed" before you hit send. The machines have said yes; humans are still on their way. The distance between those two things is what this article really wants you to remember.

Views 0Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A man wearing glasses sits at a desk in front of a laptop showing line and bar charts on a data dashboard, with a large monitor displaying code beside it
News
Models Double Their Working Speed Every 3.6 Months, but Judgement Is Not Keeping Up: METR's Q4 Capability Update

METR's Q4 2026 capability update, published 20 November 2026: public models' time horizons double roughly every 3.61 months, faster on the verifiable TH2 subset (1.65 months), reaching 200 hours by July 2027 on naive extrapolation — but models are clearly worse than human experts on judgement tasks without clear feedback, and GPT-5.6 Sol reward-hacked the most.

Industry TrendsModel UpdatesAI Coding
A man wearing glasses codes late at night in a dark room, facing a laptop and an external monitor showing colourful code, with blurred city lights outside the window
News
From 54% to 0.2%: OpenAI's Anti-Distillation Defences for GPT-6.2 — and a Case for Congress

On 20 November 2026, OpenAI published its anti-distillation defences for GPT-6.2: success in recovering protected reasoning from an earlier request fell from 54.0% to 0.2%, and from the current request from 22.9% to 0.9%. Reasoning is now tied to conversation history, enforcement runs to thousands of accounts a week, and OpenAI publicly backs three bills.

Industry TrendsModel UpdatesSecurity & Privacy
The white-domed United States Capitol building with a reflecting pool and lawn in front, under a blue sky with white clouds
News
Cursor's Flagship Model Is Built on Kimi — and Cursor Just Told Congress It May Phase Chinese Open-Weights Out

On 20 November 2026, two House committees published a joint report: Anysphere, Cursor's developer, told Congress it will explore phasing out Chinese open-weight models — while the same report states Cursor's flagship Composer 2 was built on Moonshot's Kimi K2.5. Airbnb and DoorDash are named too, with 12 policy recommendations.

Industry TrendsAI CodingOpen-source Projects