MIT Checks a Claimed Riemann Hypothesis Proof as 175,187 Lean Files Compile Without Errors

Share this article

Featured image: Bernhard Riemann in 1863 (file photo). Photo: unidentified photographer, Smithsonian Institution Libraries, Wikimedia Commons, public domain.

On Friday, Nov. 20, 2026, Anthropic mathematician Levent Alpöge claimed that he and the company’s Claude model had found a proof of the Riemann hypothesis, the 167-year-old conjecture about prime numbers that remains an unsolved Millennium Prize Problem at the Clay Mathematics Institute in Cambridge. Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, has begun independently checking the formal Lean code Alpöge posted.

Advertisement
Claim posted
Nov. 20, 2026
Alpöge’s note and Lean archive
Lean modules checked
175,187
zero compile errors at MIT
MIT checks still running
2 of 12
logged 5:05 p.m. ET; unchanged at 6:20 p.m. ET
Clay Millennium Prize
$1 million
not awarded; problem still listed as unsolved

What Alpöge posted

In a note on his website, Alpöge wrote, “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” He said the method “easily extends much further, resolving the so-called Generalized Riemann Hypothesis,” and that he and Claude had “formalized the arguments without assumptions.” He also wrote that he expected the Extended Riemann Hypothesis would follow “with more time.”

He posted the formalization as a 14-part Lean archive. Sutherland’s verification page records the joined source as 239,275,554 bytes, about 228 megabytes, a Boston Report conversion from that figure. Alpöge wrote that he was not posting a long conventional write-up. “The argument is incredibly long at the moment,” he said, adding that he intends “to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly.”

In a footnote, Alpöge wrote that the work came together for him “late Sunday night” and that he messaged Anthropic chief executive Dario Amodei about whether he could announce it. The Verge reported that he indicated he held off publication until after the company’s initial public offering on Thursday, Nov. 19. The Verge also reported that Anthropic had not commented on the announcement and had not involved itself in it. In August, the company had said it was working on the Riemann hypothesis with Claude but did not expect those techniques to prove it, The Verge reported.

Alpöge’s website identifies him as an Anthropic employee and lists a junior fellowship at the Harvard Society of Fellows. The site also posts his Harvard senior thesis, “The average elliptic curve has few integral points.” Anthropic is also a local business story: we reported Friday that Boston’s Spark Capital holds a large Anthropic stake, with a cut flowing to Massachusetts pensions.

What the Riemann hypothesis says

Bernhard Riemann stated the conjecture in an 1859 paper. It concerns the Riemann zeta function, built from the series 1 + 1/2s + 1/3s + 1/4s + … and extended across the complex plane. Some of the points where the function equals zero are well understood: the so-called trivial zeros, at the negative even integers. The Riemann hypothesis says that every other zero, called a nontrivial zero, has real part equal to 1/2. Those zeros are closely tied to how prime numbers are distributed.

Riemann showed that the nontrivial zeros lie in the critical strip between 0 and 1. In 1896, Jacques Hadamard and Charles Jean de la Vallée Poussin proved that there are no zeros on the line where the real part equals 1, a fact used to prove the prime number theorem, Quanta Magazine noted in its October account of later work. A proof of the full hypothesis would tightly constrain how far the count of primes can drift from that average.

The Clay Mathematics Institute, which is based in Cambridge, named the problem one of seven Millennium Prize Problems in 2000 and allocated $1 million to a solution of each. On its official problem page, the institute still describes the conjecture in the present tense and says the claim “has been checked for the first 10,000,000,000,000 solutions.” A later, peer-reviewed verification by Dave Platt and Tim Trudgian, posted in 2020 and published in 2021 in the Bulletin of the London Mathematical Society, proved that every zero with imaginary part up to 3,000,175,332,800 — the lowest 12,363,153,437,138 nontrivial zeros — lies on the critical line. That count is 23.6% higher than the 10 trillion figure on Clay’s page, a Boston Report calculation from the two figures. Checking any finite list of zeros cannot prove the hypothesis, because there are infinitely many of them.

Clay’s prize rules require a complete solution to be published in a qualifying outlet, to receive general acceptance in the global mathematics community, and to have been published for at least two years before the institute will consider awarding the prize. When rechecked at 6:20 p.m. ET Friday, the institute’s Millennium Prize Problems page still listed the Riemann hypothesis among its unsolved problems. It has not announced that it has accepted a proof.

What the MIT checks show

Sutherland published a running log of machine checks at math.mit.edu/~drew/grh/. The page was last updated at 5:05 p.m. ET on Friday, Nov. 20, and was unchanged when rechecked at 6:20 p.m. ET. It says the published Lean source compiled without errors in Lean 4.33.1 with Mathlib 4.33.1, that all 175,187 modules compiled, and that the five headline theorems depend only on Lean’s three standard axioms: propext, Classical.choice and Quot.sound. He also reported that Mathlib files matched the official cache, that a Lean kernel replay of every module passed, and that two independently written kernels, nanoda and con-leche, accepted the full exported proof. A con-leche comparison, he wrote, confirmed all five theorems against a separately exported “challenge” of the formal statements. The checker lean4lean accepted every compiled module; five modules passed only after a default step limit was raised.

The headline theorem in the submission, Sutherland wrote, is Mathlib’s own definition of the Riemann hypothesis: that if the zeta function vanishes at a complex number that is not 1 and is not a negative even integer, then the real part of that number is 1/2. Four further theorems include the analogous statement for all Dirichlet L-functions. On Sutherland’s page, “GRH” refers to Dirichlet L-functions only.

Two checks were still running at the time of the update: the Lean default and “paranoid” kernels on the full exported proof, and an end-to-end lake comparator run. Sutherland noted that statement comparison and the axiom check had already passed in one comparator run. He also listed caveats. The build ran on cloud machines without a sandbox. A passing check, he wrote, “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.”

Advertisement

Scientific American quoted Sutherland as saying, “The Lean proof has been compiled and independently rechecked by the Lean kernel with only the three standard axioms,” and, “Honestly the chance that anything is fishy here is approximately zero.” In an earlier Scientific American report on Friday, he called the moment historic and said, “There is much work to be done to understand the mathematical details of the proof.” Kevin Buzzard of Imperial College London, also quoted by Scientific American, said the first question when judging Lean code paired with a paper is whether the theorem stated in the code corresponds to the actual theorem, and wrote that researchers “still await a successful comparator run” to check that the code does not redefine symbols in the statement and that it compiles with several independent kernels.

By the numbers: zeros checked, and a weaker autumn result

Of the 12 substantive checks on Sutherland’s table, 10 had passed and 2 were still running as of the 5:05 p.m. ET update, or 83%, a Boston Report calculation from that table. Those totals were unchanged when the page was rechecked at 6:20 p.m. ET. Those machine checks are not the same thing as an award from the Clay Institute, a journal publication, or a proof the mathematics community has read and accepted in ordinary language.

Horizontal bar chart of 12 machine checks of a claimed Riemann hypothesis Lean proof logged by MIT’s Andrew Sutherland as of 5:05 p.m. ET on Nov. 20, 2026: 10 passed, 2 still running
Machine checks of the claimed Lean proof, as logged by Andrew Sutherland of MIT as of 5:05 p.m. ET, Nov. 20, 2026. The page was unchanged when rechecked at 6:20 p.m. ET. Source: math.mit.edu/~drew/grh/; chart: The Boston Report.

A separate, weaker result arrived in October. On Oct. 6, 2026, OpenAI released hundreds of AI-generated math papers. Among them, Quanta Magazine reported, was a Sept. 30 paper that claimed to prove a “quasi-Riemann hypothesis”: that every nontrivial zero has real part between 1/8 and 7/8, a statement formally verified in Lean. A second, shorter paper, not yet formally verified as of Quanta’s Oct. 19 article, claimed the weaker interval from 1/12 to 11/12. Science reported on Oct. 13 that the work showed there are no zeros between 7/8 and 1. James Maynard of the University of Oxford told Science that, within analytic number theory, it was “obviously the best headline result for the past 100 years in the field.” That October work is not a proof of the Riemann hypothesis, which would require every nontrivial zero to lie exactly at real part 1/2.

Bar chart showing the width of the real-part interval that can still contain nontrivial zeros of the Riemann zeta function: width 1 in 1859, width 0.75 in OpenAI’s Sept. 30, 2026 quasi-Riemann hypothesis Lean paper, and width 0 if the Riemann hypothesis is true
Width of the real-part interval in which nontrivial zeros of the zeta function may still lie. The 2026 bar is the interval 1/8 to 7/8 from OpenAI’s Sept. 30 Lean paper, as reported by Quanta Magazine on Oct. 19, 2026; it is not a proof of the Riemann hypothesis. Sources: Quanta Magazine; Clay Mathematics Institute; chart: The Boston Report.

Will Sawin of Princeton University told Quanta that once the new October methods were fully absorbed, “we will still feel as far off from the full Riemann hypothesis as we were before,” though he added, “But I am not certain of this.”

What remains unsettled

The Riemann hypothesis is still listed as unsolved by the Clay Mathematics Institute. Alpöge’s Lean archive, as of Friday, had been posted on his website rather than in a qualifying journal. Under Clay’s rules, even a proof that later wins general acceptance would have to appear in a qualifying outlet and then wait at least two years before the $1 million prize could be considered.

When the page was rechecked at 6:20 p.m. ET, those two checks were still listed as running. His log is written as a running report and may be updated after this article is published. Lean compilation and kernel checks, if they continue to hold, would be strong evidence that the formal statements as written follow from the standard axioms. They would not, by themselves, replace a human-readable account of the argument. Alpöge wrote that he has “just taken a step” and that the argument still needs to be understood and written properly. Sutherland, likewise, said there is much work to do on the mathematical details.

Sources

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up