
Photo: unidentified photographer, 1863, Smithsonian Institution Libraries (Dibner Library), via Wikimedia Commons (no known copyright restrictions)
The most famous unsolved problem in mathematics may have just fallen, and the claim arrived in the strangest form imaginable. On Friday, Nov. 20, 2026, a four-page note appeared on the personal website of Levent Alpöge, a number theorist who works at the AI company Anthropic. "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis," it begins. The note says the same method also proves the Generalized Riemann Hypothesis, a far broader statement, and that the whole argument has been checked by computer in the proof language Lean. Instead of a paper, the website offers the Lean code itself: 14 compressed archive files that unpack, according to an independent mirror, into more than 175,000 source files and about 17.4 million lines.
The short answer: Nobody knows yet whether the proof is right. If the Lean code compiles and its final theorem really states the Riemann Hypothesis, with no hidden assumptions, the 167-year-old problem would be settled in a formal sense even before any human understands the argument, and the result would rank among the biggest moments in the history of mathematics and of artificial intelligence. If the code fails, or proves something subtly different, this becomes the latest in a long line of famous claims that did not survive scrutiny. On Friday morning the first outside checks had found no obvious tricks in the files, but a full build had not finished, no expert had vouched for the mathematics, and Alpöge had not yet published a human-readable write-up.
This is a long read, because the story needs room. Below we explain what was posted and what it says, what the Riemann Hypothesis actually claims and why prime numbers depend on it, how a 33-year-old German mathematician wrote it down in 1859, and how the best minds of the next 167 years chipped at it without breaking it. Then we cover the remarkable year of AI mathematics that led to Friday’s note, what the note says about how the claimed proof works, how a Lean proof gets checked and what could still go wrong, what a proof would change, who could collect the $1 million prize, and what to watch for next.

Image: Empetrisor / Wikimedia Commons, CC BY-SA 4.0
What appeared on Friday morning
The note is titled simply "Remark." It is written in the first person and carries no byline, but it sits on alpo.ge, the site Alpöge uses for his work, and both Scientific American and New Scientist attributed it to him in stories published Friday. Scientific American broke the news on X at 6:16 a.m. Pacific time with the words "BREAKING: A mathematician at Anthropic claims to have used the company’s AI to solve the Riemann hypothesis, the most notorious open problem in all of math." As of Nov. 20, that post had been viewed about 1.3 million times, and "Riemann Hypothesis" became a trending topic in the United States. Posts on X mentioning the hypothesis or Alpöge’s name jumped from a handful an hour to about 300 in the 6 a.m. hour and more than 2,300 in the first 43 minutes after 7 a.m. Pacific. As of Friday morning, Alpöge himself had not posted about it on his X account.

Screenshot: X / @sciam
The first page is short and emotional. The Riemann Hypothesis, it says, "asks for optimal control on the count of prime numbers of a given size. Its shadows have guided mathematicians for 150 years." It then makes the claim, adds that "the method easily extends much further, resolving the so-called Generalized Riemann Hypothesis," and predicts that "with more time," the Extended Riemann Hypothesis will follow too. "Claude and I have also formalized the arguments without assumptions," it says.
It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.
The author describes working on the problem as a lifelong ambition. "It was my secret in my student days that my plans were for this problem in decades," the note says, and it credits Claude, Anthropic’s AI model, with the decisive late steps.
Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line.
— Levent Alpöge
Why only computer code? The note says the argument "is incredibly long at the moment." A footnote says the informal, AI-assisted drafts ran past 250,000 pages. Rather than post what the author calls "some massive unreadable travesty of ‘slop,’" the note says, "I decided to only present Lean." A proper write-up is promised later: "I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly."

Screenshot: X / @thomasfbloom
The timing. The second page opens bluntly: "The precise timing is horrible." Anthropic’s shares began trading on the Nasdaq on Thursday, Nov. 19, the day after the company priced the largest initial public offering ever at $83 a share. The note says the work came together "late Sunday night," that the author messaged "Dario" to ask about announcing, and that "there was a constraint to respect." Anthropic’s chief executive is Dario Amodei, though the note does not give a last name. "There is an obvious cynical view, to do with the timing, but it is false," it says, and the author takes responsibility for going public now: "I am clearly unilaterally responsible for feeling I just could not live if I waited even longer." Anthropic had not commented publicly on the claim by Friday morning.
The thanks. The note praises a long list of people, including the Princeton number theorists Manjul Bhargava, who supervised Alpöge’s doctorate, and Peter Sarnak, the analytic number theorist Maksym Radziwiłł, and several Anthropic colleagues, among them Jarred Sumner and Ralph Furman, who worked on an earlier Claude result about the same problem. Being thanked is not the same as checking a proof, and none of those named had publicly endorsed it by Friday morning. The note also, unusually, praises researchers at rival OpenAI for "good taste" in the problems they chose.
The rough timeline. The last two pages give a 20-step list that the note says Claude compiled "after a thorough search," dated from Oct. 6 to Nov. 16, 2026. It reads like a lab notebook written in the private shorthand of modern number theory. The first entry, dated Oct. 6, is "Stepanov determinant." Several key steps are dated Nov. 15. The last, dated Nov. 16, is a single acronym: "GRH."
The first outside look. Within hours, someone mirrored the 14 archive parts on the AI hosting site Hugging Face "so the artifact survives and can be checked reproducibly," with a prominent warning that the work "has NOT been verified." The mirror’s notes give the first concrete numbers about the formalization: the combined archive is 239,275,554 bytes, or about 228 megabytes, and unpacks to roughly 1.3 gigabytes, including 175,187 Lean source files with 17,374,215 lines, built against a pinned version of Mathlib, Lean’s community math library, with Lean version 4.33.1. A scan of the files found zero uses of "sorry," the Lean keyword for an unfinished step, zero new axioms, and none of several other escape hatches that can be used to fake a Lean proof. The main theorem is a single line, theorem rhOfAll : RiemannHypothesis. But the mirror’s own status line was the important one: compiling the files that the main theorem depends on was "in progress," and "a clean lake build plus #print axioms rhOfAll is what would actually settle the claim."

Screenshot: X / @thomasfbloom
The first expert reaction. "Let’s see if it compiles first," Kevin Buzzard of Imperial College London, a leading figure in Lean mathematics, told New Scientist. "Might be a while before we find the relevant experts who are able to tell us the thrust of the argument, assuming it’s all true, of course, which it probably will be because of course Alpöge will have [a] huge amount of egg on [his] face if the Lean code doesn’t compile or is deficient in some way."
A second claim, the same day. Friday’s news did not stop at one Millennium Prize Problem. The tech site OfficeChai reported that on the same day, Alpöge and the New York University mathematician Tristan Buckmaster posted a 79-page paper claiming that the Navier–Stokes equations for fluid flow can blow up in finite time with no outside force pushing the fluid, the tougher, unforced version of the problem that OpenAI’s September result left open. That claim depends on a computer "certificate" whose cryptographic fingerprint was published but whose file had not been released, so it could not yet be checked either. If both survive, Friday would go down as the day two of the Clay Mathematics Institute‘s seven $1 million problems fell.

Screenshot: X / @wtgowers
The Riemann Hypothesis, in plain language
At its heart, the Riemann Hypothesis is about prime numbers, the whole numbers greater than 1 that can only be divided evenly by 1 and themselves: 2, 3, 5, 7, 11, 13 and so on forever. Every whole number is a product of primes in exactly one way, which is why mathematicians call them the atoms of arithmetic. They are also infuriatingly irregular. There are four primes below 10, 25 below 100 and 168 below 1,000, but no simple formula tells you where the next one will be. Twin primes like 41 and 43 sit side by side, while long deserts open up elsewhere.
Yet zoom out and a pattern appears. The primes thin out at a steady, predictable rate. Around the number x, roughly one number in every ln x is prime, where ln is the natural logarithm. Near 1,000, about one number in seven is prime. Near a billion, it is about one in 21. Near a trillion, about one in 28. The Prime Number Theorem, proved in 1896, makes that precise: the count of primes up to x, written π(x), grows like x/ln x. An even better estimate, which the young Carl Friedrich Gauss proposed in the 1790s, is a smooth curve called the logarithmic integral, Li(x), which adds up the "probability" 1/ln t of each number t being prime.
The question the Riemann Hypothesis answers is how far the real count can stray from that smooth estimate. The answer it gives is: almost as little as it possibly could. If the hypothesis is true, the error is never much bigger than the square root of x, give or take a logarithm. In other words, the primes are as close to "randomly scattered at exactly the predicted density" as they can be. Here is what that looks like in practice.
| Up to x | Actual primes, π(x) | Li(x) estimate is off by | x/ln x estimate is off by |
|---|---|---|---|
| 1,000 | 168 | about 10 | about 23 |
| 1,000,000 | 78,498 | about 130 | about 6,116 |
| 1,000,000,000 | 50,847,534 | about 1,701 | about 2.6 million |
| 10,000,000,000,000 (10 trillion) | 346,065,536,839 | about 108,971 | about 12 billion |
At 10 trillion, Gauss’s estimate misses the true count of 346 billion primes by about 109,000, an error of roughly three hundred-thousandths of 1 percent, and the square root of 10 trillion is about 3.2 million. The Riemann Hypothesis says, in effect, that this kind of accuracy holds forever, no matter how far out you go. In 1901 the Swedish mathematician Helge von Koch proved that the two statements are equivalent: the error stays on the order of the square root of x times log x exactly when the hypothesis is true.
Where the zeta function comes in
So why is a statement about prime counts named after a function? The bridge was built by Leonhard Euler in 1737. Euler studied the infinite sum 1 + 1/2ˢ + 1/3ˢ + 1/4ˢ + …, now called the zeta function, ζ(s). He showed that the same quantity can be written as an infinite product with one factor for each prime: 1/(1 − 1/2ˢ) × 1/(1 − 1/3ˢ) × 1/(1 − 1/5ˢ) × … . The identity is a disguised version of the fact that every whole number factors into primes in only one way. Euler used it to give a new proof that there are infinitely many primes, and to show that the sum of the reciprocals of the primes, 1/2 + 1/3 + 1/5 + 1/7 + …, grows without limit.

Painting: Jakob Emanuel Handmann, 1753, Kunstmuseum Basel, via Wikimedia Commons (Public domain)
Euler only plugged in ordinary real numbers for s. Riemann’s leap, in 1859, was to let s be a complex number, a point on a two-dimensional plane with a "real" part measured left to right and an "imaginary" part measured up and down. The original sum only makes sense when the real part of s is bigger than 1, but Riemann showed how to extend the function smoothly across the whole plane, with a single exception at s = 1, where it shoots off to infinity. He also found a hidden mirror symmetry: the function’s values at s and at 1 − s are tied together by an exact formula, so the plane is reflected across the vertical line where the real part equals 1/2.

Image: Bernhard Riemann, 1859 (reprint scan) / Wikimedia Commons, Public domain
The places where the extended zeta function equals zero turn out to be the key to the primes. Some zeros are easy to find: ζ(s) is zero at every negative even number, −2, −4, −6 and so on. Mathematicians call those the trivial zeros. All the others, the nontrivial zeros, lie in a vertical strip between real part 0 and real part 1, called the critical strip. Riemann found an explicit formula that writes the prime count as Gauss’s smooth estimate plus a correction term for every nontrivial zero. Each zero adds a wave, and the waves together reproduce the jagged staircase of the primes exactly.

Image: Geek3 / Wikimedia Commons, CC BY 3.0
That is where the hypothesis bites. A zero with real part β contributes a wave whose size grows like x to the power β. Zeros on the center line, with real part exactly 1/2, add waves that grow like the square root of x. A single zero farther to the right, say at real part 0.7, would add a wave growing like x to the 0.7, which would eventually swamp the square-root error and make the primes far lumpier than they look. The Riemann Hypothesis says no such zero exists: every nontrivial zero sits exactly on the line where the real part equals 1/2, the "critical line." The first one is at height about 14.13, the next near 21.02, then 25.01, and they keep coming, more and more densely, forever.
That is the whole conjecture: a statement about where one function equals zero, which turns out to be the same as a statement about how evenly the primes are spread. In the official problem description he wrote for the Clay Mathematics Institute in 2000, the Italian mathematician Enrico Bombieri put the stakes plainly: "The failure of the Riemann hypothesis would create havoc in the distribution of prime numbers."
1859: the shy pastor’s son and his short paper
Georg Friedrich Bernhard Riemann was born on Sept. 17, 1826, in Breselenz, a village in the Kingdom of Hanover, the son of a Lutheran pastor. He was painfully shy and often ill, and he first enrolled at the University of Göttingen to study theology, as his father hoped. He switched to mathematics with his father’s blessing, spent two years in Berlin learning from Carl Jacobi, Jakob Steiner, Gotthold Eisenstein and Peter Gustav Lejeune Dirichlet, and returned to Göttingen in 1849 to finish a doctorate under Gauss himself.

Photo: Felina Katarina Schulz, 2020, via Wikimedia Commons (CC BY-SA 4.0)
His career was short and astonishing. His doctoral thesis introduced what are now called Riemann surfaces, a new geometric way of thinking about functions of a complex variable. His 1854 qualifying lecture, on the hypotheses that lie at the foundations of geometry, sketched the curved-space geometry that Albert Einstein would use six decades later for general relativity. In 1859, after Dirichlet’s death, Riemann took over the chair at Göttingen that had once been Gauss’s, and the same year the Berlin Academy of Sciences elected him a corresponding member.

Photo: Julian Herzog, 2017, via Wikimedia Commons (CC BY 4.0)
New members were expected to report on their research, and Riemann’s report, published in the academy’s monthly proceedings in November 1859, was titled "Ueber die Anzahl der Primzahlen unter einer gegebenen Grösse," or "On the Number of Primes Less Than a Given Magnitude." It runs to only about ten printed pages and is the only paper he ever published on number theory. In it he extended the zeta function to the complex plane, proved its mirror symmetry, introduced the explicit formula that ties the primes to the zeros, and stated, almost in passing, the guess that would carry his name. Writing about the zeros in a slightly different form of the function, he said:
… it is very probable that all roots are real. Of course one would wish for a rigorous proof here; I have for the time being, after some fleeting vain attempts, provisionally put aside the search for this, as it appears dispensable for the immediate objective of my investigation.
"All roots are real," in the form of the function Riemann was using, is the same as saying all the nontrivial zeros of zeta lie on the critical line. For a paper that would shape mathematics for more than a century and a half, the hypothesis itself got one sentence and a shrug.
Riemann did not live to return to it. He married Elise Koch in 1862, saw his health fail soon after, from what became tuberculosis, and spent his last years traveling to Italy in search of a healthier climate. He died at Selasca, on the shore of Lake Maggiore, on July 20, 1866, two months before his 40th birthday, and was buried in the nearby cemetery at Biganzolo. Back in Göttingen, his housekeeper threw out some of the papers in his office, including unpublished work, before relatives stepped in. What survived went to the university library in Göttingen, and in 1932 the mathematician Carl Ludwig Siegel dug through those notes and found that Riemann had been sitting on far more than he published: a sophisticated formula for computing zeta’s values, now called the Riemann–Siegel formula, and hand calculations that located the first few zeros to several decimal places. The formula became the workhorse of every serious computer search for zeros in the 20th century.

Photo: g.rondena, 2009, via Wikimedia Commons (Public domain)
The scale of the paper’s influence is hard to overstate. The 1859 memoir set the agenda for what became analytic number theory. Its methods led directly to the proof of the Prime Number Theorem 37 years later. And its one unproven sentence turned into the single most famous open question in mathematics.
167 years of near misses
What follows is a compressed history of the siege, from the first proof that the zeros behave at all to the computer runs that checked trillions of them. It explains why the problem was considered so hard, and why Friday’s claim is so startling.
1792 or 1793: Long before Riemann, the teenage Gauss studied tables of primes and noticed that their density near a number x is about 1/ln x. He recalled the observation decades later in an 1849 letter to the astronomer Johann Encke. Adrien-Marie Legendre published a similar formula in 1798. Neither could prove it.
1896: Jacques Hadamard in Paris and Charles-Jean de la Vallée Poussin in Belgium independently proved the Prime Number Theorem. The crucial step, in both proofs, was showing that zeta has no zeros on the line where the real part equals 1, the right-hand edge of the critical strip. Even that sliver of control over the zeros was enough to pin down the primes’ average density. It also showed how much more the full hypothesis would give.
1900: At the International Congress of Mathematicians in Paris, David Hilbert presented his famous list of problems for the new century. The Riemann Hypothesis was part of the eighth, alongside Goldbach’s conjecture and the twin prime conjecture. It is the only problem from Hilbert’s list that also made the Clay Institute’s Millennium list a century later. Hilbert is often quoted as saying that if he woke after sleeping for a thousand years, his first question would be whether the Riemann Hypothesis had been proved.

Photo: unknown photographer, before 1912, via Wikimedia Commons (Public domain)
1914: Two English mathematicians made the first big dents. G. H. Hardy proved that infinitely many zeros lie on the critical line, though not what fraction of them. The same year, his collaborator J. E. Littlewood proved something unsettling: Gauss’s estimate Li(x), which overcounts the primes at every value anyone had ever checked, must eventually undercount them, and the two curves cross infinitely often. In 1933 Stanley Skewes showed, assuming the Riemann Hypothesis, that the first crossing happens before a number so large it is usually written as 10 to the 10 to the 10 to the 34th power. The lesson stuck: in this subject, trillions of numerical checks are evidence, never proof.
1932: Siegel published the Riemann–Siegel formula from Riemann’s papers. Within a few years, E. C. Titchmarsh at Oxford used it, with punched-card tabulating machines and their operators, to show that the first 1,041 nontrivial zeros all lie on the line.
1942: The Norwegian mathematician Atle Selberg proved that a positive proportion of all the zeros lie on the critical line. His proportion was tiny, but it was the first time anyone could say "a fixed fraction" rather than just "infinitely many." Selberg won a Fields Medal in 1950, partly for this work.
1950 to 1953: Alan Turing, who had wondered about the zeros since before the Second World War, ran the problem on the Manchester Mark 1, one of the first stored-program computers, and published the results in 1953. He was hunting for a counterexample. "The calculations were done in an optimistic hope that a zero would be found off the critical line," he wrote. He found none, but he invented a clever way to certify that a search had not missed any zeros, still called Turing’s method and still used today.
1956 to 1968: Electronic computers took over. Derrick Lehmer used the SWAC computer to check the first 25,000 zeros. In 1966, R. Sherman Lehman used an IBM 7090 at the University of California, Berkeley, to check 250,000. By 1968, J. Barkley Rosser, J. M. Yohe and Lowell Schoenfeld had pushed the count to 3.5 million. Every zero was on the line.

Photo: NASA/Don Richey, 1963, public domain
1972: The most famous cup of tea in mathematics. Hugh Montgomery, a young number theorist visiting the Institute for Advanced Study in Princeton, had worked out, assuming the hypothesis, a formula for how pairs of zeros are spaced along the line. At afternoon tea he mentioned it to the physicist Freeman Dyson, who recognized it at once: it was the same formula that describes the spacing of the eigenvalues of large random matrices, the mathematics physicists had been using since the 1950s to model the energy levels of heavy atomic nuclei. The zeros of zeta, a pure-math object, seemed to behave like the energy levels of a quantum system. The coincidence revived an old idea credited to Hilbert and George Pólya: that the zeros might be the eigenvalues of some yet-unknown physical operator, which would force them onto the line. Nobody has found that operator.

Photo: Zeete, 2023, via Wikimedia Commons (CC BY-SA 4.0)
1974: Norman Levinson of MIT proved that more than one-third of the zeros lie on the critical line, using a new technique called mollification. The same year, Pierre Deligne finished proving the Weil conjectures, including the version of the Riemann Hypothesis for curves and higher-dimensional shapes over finite fields. More on why that matters below.
Late 1980s: Andrew Odlyzko of AT&T Bell Labs, using a much faster algorithm he developed with Arnold Schönhage, computed huge batches of zeros far up the line, eventually including zeros near the 10²⁰th. Their spacing matched the random-matrix prediction with eerie precision. In 1986, a team in Amsterdam led by Jan van de Lune checked the first 1.5 billion zeros.
1989: Brian Conrey raised Levinson’s one-third to two-fifths of all zeros.
2000: On May 24, 2000, at a meeting in Paris, the Clay Mathematics Institute announced seven Millennium Prize Problems, each with a $1 million reward. The Riemann Hypothesis was on the list, with the official description written by Bombieri.
2001 to 2005: A volunteer computing project called ZetaGrid harnessed thousands of volunteers’ computers and checked roughly 900 billion zeros. In 2004 Xavier Gourdon and Patrick Demichel reported checking the first 10 trillion zeros with a faster method. That is the figure the Clay Institute’s own page still cites.
2004: Louis de Branges, the Purdue mathematician who had famously proved the Bieberbach conjecture in 1984, posted a claimed proof of the Riemann Hypothesis. It was not accepted. Earlier, in 2000, Brian Conrey and Xian-Jin Li had shown that a positivity condition at the heart of de Branges’s approach fails.
2018: On Sept. 24, 2018, Sir Michael Atiyah, one of the most decorated mathematicians alive, presented a "simple proof" in a talk at the Heidelberg Laureate Forum. Experts quickly concluded it was not a valid proof. Atiyah died the following January. The episode is a reminder that even a towering reputation is no substitute for a checkable argument.
2018 to 2020: Brad Rodgers and Terence Tao proved that a quantity called the de Bruijn–Newman constant is at least zero, while the Riemann Hypothesis is equivalent to that constant being exactly zero. In Charles Newman’s phrase, if the hypothesis is true, it is "only barely so." In 2020, Kyle Pratt, Nicolas Robles, Alexandru Zaharescu and Dirk Zeindler raised the proven fraction of zeros on the line to five-twelfths, about 41.6 percent, where it stood until this year.
2021: David Platt and Timothy Trudgian published the most extensive rigorous check to date, confirming that every zero up to a height of about 3 trillion lies on the critical line. That range contains more than 12 trillion zeros.
2024: Larry Guth of MIT and James Maynard of Oxford improved a bound on how many zeros could possibly lie off the line, the first real progress on that question since Albert Ingham’s estimate in 1940. Experts hailed it as a breakthrough. It still left the hypothesis itself far out of reach.
Even this partial list shows the shape of the problem. Every generation produced real theorems about the zeros, each one hard-won, and each one fell short of the full statement. Before this year, most number theorists believed no existing technique could reach the hypothesis, and that a proof would need ideas nobody had yet imagined.
Why mathematicians believed it, and why some doubted
Before Friday, most mathematicians were confident the Riemann Hypothesis is true, even though nobody could prove it. That confidence rested on several independent kinds of evidence, and it is worth separating them, because each tells a different story about what a proof must look like.
The numbers. More than 12 trillion zeros have been rigorously checked, and every one lies on the critical line. Odlyzko’s spot checks went much farther, to batches of zeros near the 10²³rd, and found the same. But as Littlewood’s 1914 result showed, numerical evidence in this subject can mislead at enormous scales, because the relevant effects often grow like the logarithm of the logarithm of the numbers involved, which barely moves even across trillions.
The near misses. In 1956 Lehmer noticed pairs of zeros that sit unusually close together on the line, now called Lehmer pairs. A zero off the line would have to come with a mirror-image partner, and close pairs look like places where two zeros almost left the line together. They never quite do, and the random-matrix model predicts close pairs at exactly the observed rate. Skeptics read them as warning signs, believers as more evidence that the zeros behave like eigenvalues, which are forced to repel each other.
The statistics. Montgomery’s pair correlation, Odlyzko’s computations and later work by Jon Keating, Nina Snaith and others showed that the zeros match random-matrix predictions in more and more detail. If the zeros really are the spectrum of some hidden symmetric operator, they would have to lie on the line. This is evidence about mechanism, not just about individual cases.
The analogy. The function-field version is a theorem, proved by Weil and Deligne. When a statement is true in every analogous setting mathematicians can reach, it is natural to believe it in the one they cannot.
The partial results. Hardy, Selberg, Levinson, Conrey and the 2020 team showed that a growing fraction of the zeros lie on the line, and the 2026 results pushed that past two-thirds. Every partial result pointed the same way.
Not everyone was persuaded. Littlewood himself wrote in 1962 that he believed the hypothesis to be false and saw no imaginable reason for it to be true. Turing, as noted above, ran his computer search hoping to find a counterexample. The Serbian number theorist Aleksandar Ivić published papers arguing that the evidence was weaker than generally believed. And the de Bruijn–Newman result shows that if the hypothesis is true, it is true with no room to spare. None of the doubters found a zero off the line, but their caution is a useful reminder that belief, however widespread, is not proof.
How zeros are checked today
Checking a zero is harder than it sounds, because a computer cannot evaluate the zeta function exactly. Modern searches combine three ideas, two of which go back decades. The Riemann–Siegel formula, found in Riemann’s own notes, gives a fast way to approximate zeta on the critical line. Turing’s method certifies that a search has not skipped any zeros, by comparing the number found against an exact count derived from the function’s symmetry. And interval arithmetic, used in Platt and Trudgian’s 2021 verification, replaces each approximate number with a guaranteed range, so that rounding errors cannot produce a false result. Together they let a computer prove, not merely suggest, that every zero in a given range is on the line. They can never reach all of them, which is why only a proof can settle the question.

Photo: Elliott & Fry, 1951, via Wikimedia Commons (Public domain)
The function-field shortcut: the version that was already proved
To understand what Friday’s note claims to have done, you need to know about a strange fact: there is a version of the Riemann Hypothesis that mathematicians proved decades ago. It lives in a parallel world of arithmetic, and the note says its method is "a descendant" of one of those old proofs.
The parallel world is arithmetic with a finite number of "digits." Imagine a clock with 101 hours, where counting past 100 wraps back to zero. You can add, subtract, multiply and divide on such a clock, which mathematicians call a finite field. You can also ask how many solutions an equation has on that clock, for instance how many pairs of hours x and y satisfy the equation of a curve like y² = x³ + x + 1. Counting those solutions turns out to be the exact analog of counting primes.
In his 1921 doctoral thesis, the Austrian mathematician Emil Artin defined zeta functions for such curves and conjectured that they obey their own Riemann Hypothesis. In this world, the hypothesis has a down-to-earth meaning: the number of solutions is always very close to the expected count, off by at most a square-root-sized error. For the simplest family, elliptic curves, the German mathematician Helmut Hasse proved it in 1933. Over a clock with q hours, an elliptic curve always has a number of points within 2√q of q + 1. On the 101-hour clock, for example, that means every such curve has between 82 and 122 points, never fewer, never more.
André Weil proved the Riemann Hypothesis for every curve over a finite field. He announced the result in 1940, in a note written while he was in a French military prison for failing to report for duty, and published the complete proof in 1948. The next year he proposed a far-reaching set of conjectures, now called the Weil conjectures, that extended the statement to shapes of any dimension. Proving them took a generation and reshaped mathematics. Alexander Grothendieck rebuilt much of algebraic geometry from scratch, partly to attack them. His student Pierre Deligne finally proved the last and hardest part, the Riemann Hypothesis piece, in 1974, work that earned him a Fields Medal in 1978.

Photo: Eecc, 2005, via Wikimedia Commons (Public domain)
Stepanov’s trick. Weil’s and Deligne’s proofs use heavy machinery. In 1969 the Soviet mathematician Sergei Stepanov found a startlingly elementary route for some cases, and in 1973 Bombieri extended it to all curves. The idea is to build a cleverly chosen auxiliary polynomial that vanishes at every solution of the equation, and not just vanishes, but vanishes to high order, the way x² vanishes "twice" at zero. Mathematicians call that order of vanishing the multiplicity. Because a polynomial of a given degree can only have so many zeros counted with multiplicity, packing in lots of multiplicity puts a ceiling on the number of solutions. A symmetrical argument gives a floor. Squeeze from both sides, and you get the square-root bound.
Why it never transferred. For the ordinary whole numbers, nobody could make Stepanov’s trick, or Weil’s or Deligne’s, work. The finite-field world has features the integers lack. Most important, the integers have no natural notion of a derivative, and multiplicity is a statement about derivatives: a polynomial vanishes to order two at a point when both it and its derivative are zero there. You can differentiate a polynomial, but what would it mean to differentiate the number 17? For decades, some mathematicians have hunted for a hidden "field with one element" that would let the integers be treated like a curve over a finite field, so the old proofs could be copied over. That program has produced beautiful mathematics but no proof.
This is the gap Friday’s note claims to bridge. "In Stepanov’s argument resolving Riemann over function fields one uses a divisor involving multiplicity," it says. "This argument is a descendant of Stepanov’s, but multiplicity is hard to find over the integers, corresponding to something like differentiating whole numbers." The proposed fix, as best a non-specialist can parse it, is to get the effect of multiplicity from somewhere else entirely: from a chain of transformations called theta lifts in the theory of automorphic forms, a branch of number theory that studies functions with enormous amounts of symmetry. Each lift, the note suggests, mimics one extra order of vanishing. The note calls this the line that "will suffice to make fellow mathematicians fall out of their chairs": for a sum over primes up to a size X, Claude suggested that the chain of lifts should have length √X. "And yet, what else could it be?" the note adds.
We cannot tell you whether that idea works. Neither, on Friday, could anyone else in public. But it explains why specialists took the note seriously enough to start reading the code rather than dismissing it. It names a specific, long-standing obstacle, and a specific, technically plausible way around it, in the language of people who know the field.
2026: the year AI came for number theory
A year ago, a claim like Friday’s would have been laughed off within minutes. It was not, because of what happened in the months before it. Our earlier piece on how AI caught mathematicians off guard tells the wider story. Here is the thread that leads to the Riemann Hypothesis.
July 2025, olympiad gold. AI systems from Google DeepMind and OpenAI each reached a gold-medal score at the International Mathematical Olympiad, solving five of six problems in plain-language proofs. That was competition math, not research, but it showed language models could sustain a long, careful argument.
January to May 2026, the Erdős problems. AI systems began settling problems from the famous list left by Paul Erdős, first with asterisks and then without them. On May 20, 2026, OpenAI announced that an internal model had disproved Erdős’s unit distance conjecture with a construction drawing on deep algebraic number theory. Nine outside mathematicians checked it. The Fields medalist Tim Gowers wrote that if a human had submitted the paper to the Annals of Mathematics, he "would have recommended acceptance without any hesitation."
July 2026, the Jacobian conjecture. A version of Claude found a counterexample to the Jacobian conjecture, a question about polynomial maps that dates back to Ott-Heinrich Keller in 1939 and appears on Stephen Smale’s list of problems for the 21st century. The counterexample, prompted by a question from the mathematician Akhil Mathew, was short enough to write out in 216 characters.
Aug. 10, 2026, two-thirds of the zeros. This was the result that put Anthropic and Alpöge directly on the Riemann trail. Anthropic said an unreleased Claude research model had raised the proven fraction of zeta zeros lying on the critical line from 41.6 percent, the 2020 record, to about 67.2 percent. The company’s account said the model used about 31 million output tokens, coordinated about 60 subagents, considered some 650 ideas, ran about 2,400 shell commands and read 54 papers from the arXiv preprint server. The work began with a prompt from the Anthropic engineer Jarred Sumner. Alpöge and Ralph Furman validated the mathematics, Eric Easley led a complete formalization in Lean, and the veteran zeta-function experts Brian Conrey and Daniel Goldston examined the result. The improvement was huge by the standards of the field, where the fraction had crept up a few points per decade since Levinson. It was still not the hypothesis, which needs 100 percent, with no exceptions.
Sept. 8, 2026, Navier–Stokes. OpenAI announced that a swarm of roughly 10,000 coordinating AI agents, running for about 88 hours, had found a proof that a fluid pushed by a smooth outside force can develop a singularity in finite time, the first claimed solution of a Millennium Prize Problem since the Poincaré conjecture. A model then wrote a formal Lean version in about 17 hours. Three days later, the Clay Institute said the problem "has apparently been settled." Buckmaster and Alpöge had used AI to reach a closely related result just hours earlier, and credit has been disputed since.
September 2026, the backlash. Twenty-six Fields medalists, by the count of the Swiss university EPFL, including Bhargava, Deligne, Tao, Maynard and Peter Scholze, signed a declaration titled "A Severe Misalignment of AI in Mathematics." It argued that AI companies’ use of famous open problems as benchmarks clashes with the core aim of mathematics, conceptual understanding, and that a flood of machine-made true-or-false verdicts could undermine credit and the human transmission of ideas. Notice whose names were on it: Alpöge’s doctoral adviser, and the mathematician who proved the function-field Riemann Hypothesis.
Oct. 6, 2026, the quasi-Riemann Hypothesis. OpenAI released 377 results across 722 papers. One of them claimed what number theorists call the quasi-Riemann Hypothesis: that every nontrivial zero of zeta, and of a large family of related functions, lies strictly inside a strip narrower than the one Riemann knew. A 199-page version, dated Sept. 30 and verified in Lean, said all the zeros have real part between 1/8 and 7/8. A 49-page version, dated almost a week later and the only one of the 722 papers edited by a person, proved a slightly weaker strip, between 1/12 and 11/12, by what experts said were similar methods. Quanta Magazine reported that Peter Sarnak had repurposed his Princeton course to study it, and that experts largely agreed it was correct. Jared Duker Lichtman of Stanford called it "the most significant AI-for-math achievement so far, by quite a margin." Maksym Radziwiłł of New York University said the implications were "staggering," because "the vast majority of consequences of RH already follow from quasi-RH." The key family of functions involved objects called metaplectic forms, and the Rutgers mathematician Henryk Iwaniec had narrowly missed the same family in 2002.
Even then, few expected the full hypothesis to fall soon. "It seems most likely to me that once mathematicians have fully internalized the new methods and worked through their implications, we will still feel as far off from the full Riemann hypothesis as we were before," Will Sawin of Princeton told Quanta. Thomas Bloom of the University of Manchester agreed that the full statement still seemed out of reach, then added: "we would have said the same about quasi-RH a few weeks ago! At this point nothing would surprise me."
October 2026, money for verification. The Simons Foundation launched a $15 million effort to help mathematicians evaluate, understand and build on AI-generated results, with $5 million of it from OpenAI. OpenAI also set up an independent advisory group on mathematics and AI, hosted by the Institute for Advanced Study, to advise on how such results are reviewed and communicated.
Nov. 17 to 19, 2026, Anthropic goes public. Anthropic published another research post, this one on a physics calculation about black holes. The next day it priced the largest initial public offering ever at $83 a share, and on Nov. 19 its stock began trading on the Nasdaq under the ticker ANT.
Nov. 20, 2026, the claim. One day later, an Anthropic mathematician posted a note claiming the Riemann Hypothesis itself.
Read in sequence, the note looks less like a bolt from the blue and more like the next step on a staircase. It even says so. It praises OpenAI’s quasi-Riemann work, calling one step "strictly dominated" by another trick in the same release, and says the author recognized it as "Stepanov without multiplicity." That is a striking remark. If it is right, the OpenAI proof of the weaker statement and Friday’s claimed proof of the full one are cousins: one uses the Stepanov idea with the multiplicity left out, the other claims to have found a way to put it back in.
Inside the claimed proof, as far as anyone can tell
A warning before we start: what follows is a reading of a four-page note whose last two pages are written in dense, private shorthand, and which itself says parts of its summary "might make zero sense." No human-readable proof has been released. Specialists will spend months, at least, working out what the Lean code actually does. With that said, the note gives enough clues to sketch the shape of the argument and to explain why its ingredients caught experts’ attention.
The starting point: counting primes with weights
Number theorists rarely count primes directly. Instead they use the von Mangoldt function, written Λ(n), which equals the logarithm of p when n is a prime p or a power of one, and zero otherwise. Adding up Λ(n) for every n up to X gives a weighted prime count that is about X, and the Riemann Hypothesis is equivalent to that sum never straying from X by much more than the square root of X. Step 5 of the note’s timeline, dated Oct. 20, records the decision to "definitively switch" to a smoothed version of this sum, with the main terms subtracted off, and mentions the Guinand–Weil explicit formula, a modern descendant of Riemann’s formula linking primes and zeros. The note says that formula "had the sign," meaning a positivity property the argument needed.
That word, sign, is the thread through the whole note. Many approaches to the hypothesis come down to showing that some carefully built quantity can never be negative. If you can prove the right quantity is always at least zero, a zero off the critical line would force it negative, and the contradiction rules such zeros out. The trick is finding a quantity you can actually control. Steps 4 and 6 of the timeline describe exactly this kind of positivity statement: "W(1S) ≥ 0 via a determinant," and a later inequality that "gives > 1 − η," which reads like a step that gets arbitrarily close to the full result.
The toy model: a new proof of an old theorem
The note says the project began by "combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph." The second page names that shadow more specifically: the author says he had found a "sign-only proof of Weil and Deligne," built around an identity written qN = Mᵏ + R. In plain terms, the claim seems to be a new and much simpler proof of the finite-field Riemann Hypothesis, the version Weil and Deligne proved, using only a sign argument: two sides of an equation that must have opposite signs unless the error term is small. If that alone holds up, it would be a significant result. The note mentions it almost in passing.
The leap: multiplicity from symmetry
From there, the argument tries to carry the same structure over to the ordinary integers, where multiplicity, the key ingredient in Stepanov’s method, has no obvious meaning. The note’s answer comes from automorphic forms, functions so symmetric that they behave like the harmonics of a vibrating drum shaped by number theory. Those functions come attached to symmetry groups, and a tool called the theta correspondence, or theta lifting, transports them from one group to a bigger one.
According to the timeline, the decisive breakthrough was finding "multiplicity 2" on a group called O(2, n) through what the note calls a "natively-automorphic" determinant, and then seeing that each further theta lift could win one more order of multiplicity. Step 8, dated Nov. 15, describes "iterated theta lifts" through a chain of unitary groups that "wins multiplicity m," with the note adding: "Time to take m like √X." In other words, to control a prime sum of length X, the argument stacks up about √X lifts, each one contributing a sliver of the vanishing that a derivative would give over a finite field.
The note is candid that this is not the author’s home turf. "I personally have little mastery of automorphic forms on higher rank groups (youthful foolish stubbornness)," it says, and that is precisely where it credits Claude’s "technical mastery."
A parts list from across number theory
What makes the timeline so striking to specialists is the sheer range of machinery it names. Among the 20 steps:
- Gan–Gross–Prasad periods (step 7, Oct. 14), special integrals of automorphic forms whose values are tied to L-functions by a celebrated set of conjectures, much of which has been proved in recent years.
- A relative trace formula (step 16, Oct. 14), one of the heaviest tools in the modern theory, used here, the note says, together with a bound on "a suitably projected Poincaré series."
- The seesaw principle and Rallis (step 12, Nov. 15), a classic identity in theta-lift theory, here used to show that a period of the iterated lift does not depend on the particular characters chosen.
- The Chowla–Selberg formula, Weyl and the Siegel mass formula (step 9, Nov. 15), used, the note says, for "cancellation of order-m terms with proven divisibility."
- An Iwasawa tower (steps 11 and 19), a tower of ever-larger number fields that the note says plays the role of the tower of finite fields in the function-field proof, chosen "specifically using Behrend!!", a nod to Felix Behrend’s 1946 construction of large sets of whole numbers with no three in arithmetic progression. A footnote says this trick is "obviously going to be unnecessary."
- The Bost slope inequality (step 13, Oct. 6), a tool from Arakelov geometry, the branch that treats whole numbers geometrically.
- A "significant weakening of Vandiver" (step 17, Nov. 15). Vandiver’s conjecture is a century-old open question about the arithmetic of cyclotomic fields, the number systems built from roots of unity. The note says the argument needed to "fight for" a weak form of it, and a footnote adds that what it proved "is still pretty strong!!" and may imply that certain algebraic invariants called K-groups of the integers vanish.
Two things stand out even to a non-specialist reading this list. First, if the argument is right, it does not merely settle the hypothesis; it seems to drag along new results in several other areas, which is either a sign of a deep proof or of a long chain with many places to break. Second, the components are mostly known, powerful theories assembled in an unexpected order. That matches what Quanta reported about OpenAI’s quasi-Riemann proof, which experts described as "a surprising synthesis of well-known ideas" rather than a brand-new theory.
The Generalized Riemann Hypothesis
The note’s final step, dated Nov. 16, is "GRH." The Generalized Riemann Hypothesis makes the same claim, every nontrivial zero on the critical line, for an infinite family of related functions called Dirichlet L-functions, which govern how primes split among arithmetic progressions, such as primes ending in 1, 3, 7 or 9. The Lean entry point found by the Hugging Face mirror is consistent with that: the main theorem is derived from a statement covering every character, the technical objects that index Dirichlet L-functions, of which the ordinary zeta function is the simplest case. The note says it expects the Extended Riemann Hypothesis, the corresponding statement for still more general number systems, to follow "with more time."
There is also a moment of comic relief. A footnote says Claude "cutely called it the GRH… ‘book,’" and that its Theorem 1 began by observing that, because real numbers are totally ordered, a number that is not 1/2 must be either bigger or smaller than 1/2. It is a reminder that machine-written mathematics can be both astonishing and absurdly literal, and why the author chose not to post the prose.
How a Lean proof gets checked, and what could still go wrong
The single most important fact about Friday’s claim is that it does not ask anyone to trust a person or a company. It asks them to run a program. That is what makes it unlike every previous claimed proof of the Riemann Hypothesis, and it is also where the real tests lie.
What Lean is
Lean is a programming language and proof assistant created by Leonardo de Moura, who started it at Microsoft Research in 2013. Writing a proof in Lean means expressing every definition and every logical step in a formal language that a small, carefully audited core program, called the kernel, can check mechanically. If the kernel accepts a proof, the only things you need to trust are the kernel itself, a handful of standard logical axioms, and the statement of the theorem. Over the past decade, Lean’s community library, Mathlib, has grown into a vast shared encyclopedia of formalized mathematics, with more than 210,000 theorems as of May 2025, and it includes the definition of the Riemann zeta function and of the hypothesis.
Here is the official Mathlib definition, the target any formal proof must hit:
def RiemannHypothesis : Prop := ∀ (s : ℂ) (_ : riemannZeta s = 0) (_ : ¬∃ n : ℕ, s = -2 * (n + 1)) (_ : s ≠ 1), s.re = 1 / 2
In words: for every complex number s, if zeta of s is zero, and s is not one of the trivial zeros −2, −4, −6 and so on, and s is not 1, then the real part of s is 1/2. The archive’s main theorem, theorem rhOfAll : RiemannHypothesis, claims exactly that statement, if, and this is a big if, the name RiemannHypothesis in the archive means the same thing it means in Mathlib.
The four questions checkers will ask
1. Does it compile? The most basic test is whether the entire project builds with Lean’s standard build tool, lake, with no errors. With 175,187 files and 17.4 million lines, that is a serious computing job. The mirror reports that the project raises Lean’s per-step resource limit, maxHeartbeats, to 4 million, 20 times the default of 200,000, which suggests some steps are very heavy. On Friday morning the build was still running.

Photo: NASA/Dominic Hart, 2008, public domain
2. What does it depend on? Lean has a command, #print axioms, that lists every axiom a theorem ultimately relies on. A clean answer would be only Lean’s three standard axioms: propositional extensionality, the axiom of choice (in Lean, Classical.choice) and quotient soundness. Anything else, especially a custom axiom or the use of "sorry," would mean the proof assumes something it has not proved. The mirror’s text scan found no new axioms and no sorries, but a text scan is not the same as asking Lean, because Lean code can generate declarations through metaprogramming that never appear as plain text. The mirror counted 32 uses of run_tac and 30 of liftMetaTactic, two of the commands that let Lean code write or run other Lean code. Those are often harmless, but checkers will want to look at every one.
3. Is the statement the real one? A formal proof is only as good as its statement. The archive is built against a pinned version of Mathlib, and checkers will compare its definitions of riemannZeta and RiemannHypothesis, character by character, against the official library. A subtly altered definition, for example a zeta function defined differently or a hypothesis that quietly excludes some zeros, could produce a perfectly valid proof of the wrong theorem. Mathematicians call this the statement-fidelity problem, and it is why Lean experts say the theorem statement is the part humans must always read themselves.
4. Does an independent checker agree? Lean’s kernel is small, but it is still software. The gold standard is to export the finished proof and replay it in an independent checker written by different people. When OpenAI released its quasi-Riemann proof in October, outside researchers replayed its Lean statements in a second kernel, including one called nanoda, and reported that both accepted them. The same will happen here, on a far larger scale.
If the answers are yes, clean, yes and yes, the mathematical community will be in a position it has never been in before: holding a machine-verified proof of the Riemann Hypothesis that no human yet understands.
Precedents, and why this one is different
Formal verification of landmark results is not new. Georges Gonthier and colleagues formalized the four color theorem in the Coq proof assistant, completing the work in 2005. Thomas Hales’s Flyspeck project finished a formal proof of the Kepler conjecture on sphere packing in 2014, after a human referee team had said it was only "99 percent certain" of his original computer-assisted proof. In 2020, Peter Scholze challenged the Lean community to verify a key theorem in his new theory of condensed mathematics, and the Liquid Tensor Experiment, led by Johan Commelin, completed that challenge in 2022. Buzzard is leading a multi-year project to formalize the proof of Fermat’s Last Theorem in Lean.
In every one of those cases, the human proof came first and the formal proof confirmed it. This year inverted the order. OpenAI’s Navier–Stokes and quasi-Riemann results, and Anthropic’s two-thirds result, arrived with Lean proofs alongside, or ahead of, human understanding. Friday’s claim pushes that to the extreme: Lean only, with the human explanation promised later. Even if the code checks out, mathematicians will not consider the work finished until they understand why it is true. As Sarnak said of the quasi-Riemann proof, he would not be satisfied it is "a Theorem" until humans check and understand all the details.
What a proof would change
If the Riemann Hypothesis is true, mathematicians have in a sense been living with its consequences all along: the primes have always been as regular as it says. What changes with a proof is certainty. Thousands of published theorems begin with the words "assuming the Riemann Hypothesis" or "assuming GRH," and a proof would turn every one of them into an unconditional fact overnight. Here are some of the most concrete.
Sharper prime counts
The hypothesis gives an explicit, guaranteed error bar for the prime count. In 1976 Lowell Schoenfeld showed that, assuming it, the difference between π(x) and Li(x) is less than √x × ln x divided by 8π for every x of at least 2,657. Without the hypothesis, the best proven error bounds are vastly weaker. A proof would make the table near the top of this article into a theorem for every row that will ever be computed.

Painting: Christian Albrecht Jensen, 1840, via Wikimedia Commons (Public domain)
Gaps between primes
Primes get sparser as numbers grow, but how long can a gap between consecutive primes be? In 1920 the Swedish mathematician Harald Cramér showed that, under the Riemann Hypothesis, the gap after a prime p is at most on the order of √p × ln p. The best unconditional result, proved by Roger Baker, Glyn Harman and János Pintz in 2001, only guarantees a prime in every interval of length about x to the power 0.525 near x. Neither comes close to what most experts believe, which is that gaps are no bigger than about (ln p)², so even a proof would leave plenty of open questions about primes.
Faster, certain primality tests
Computers need to know whether huge numbers are prime, for example when generating encryption keys. The fast Miller–Rabin test is randomized and could in principle be fooled. In 1976 Gary Miller showed that, if the Generalized Riemann Hypothesis is true, checking a small, explicit list of bases is enough to make the test deterministic, and in 1990 Eric Bach showed that it suffices to check every base below 2(ln n)². A proof of GRH would make that fast deterministic test unconditionally correct. In 2002, Manindra Agrawal, Neeraj Kayal and Nitin Saxena of the Indian Institute of Technology Kanpur found a deterministic polynomial-time primality test that needs no unproved hypothesis, but the GRH-based approach is far faster in practice.
Old problems with GRH-shaped holes
Many famous results were first proved conditionally. In 1923 Hardy and Littlewood showed that the Generalized Riemann Hypothesis would imply that every large enough odd number is the sum of three primes. In 1997 a team led by Jean-Marc Deshouillers showed that GRH implies the statement for every odd number above 5, and in 2013 Harald Helfgott finally proved it without the hypothesis. In 1967 Christopher Hooley showed that GRH implies Artin’s conjecture on primitive roots, which predicts, for example, that 2 is a primitive root for about 37.4 percent of primes. In 1952 Nesmith Ankeny showed that GRH implies the smallest number that is not a perfect square modulo a prime p is at most a constant times (ln p)². A proof of GRH would make all of these, and many more, settled facts.
Strange equivalent statements
The hypothesis also has disguises that look nothing like zeta zeros. In 1984 the French mathematician Guy Robin proved that it is true if and only if the sum of the divisors of n is less than e^γ × n × ln ln n for every n greater than 5,040, where γ is Euler’s constant, about 0.5772. In 2002 Jeffrey Lagarias turned that into an elementary statement involving nothing more than harmonic numbers, the sums 1 + 1/2 + 1/3 + … + 1/n. A proof of the hypothesis proves all of these at once.
What it would not do
It would not break encryption. This is the most common myth about the problem. Widely used encryption schemes such as RSA rely on the difficulty of factoring the product of two huge primes. A proof that the primes are regularly distributed on average says nothing about how to find the factors of a particular number. Nothing in the note, or in the theory, suggests a factoring algorithm.
It would not change your phone, bank or computer. Engineers and cryptographers already behave as if the hypothesis is true. The practical impact of a proof would be almost entirely inside mathematics, where it would be enormous.
It would not settle everything about primes. The twin prime conjecture, Goldbach’s conjecture for even numbers and the question of whether there is always a prime between two consecutive perfect squares would all remain open. Hilbert grouped some of them with Riemann in his eighth problem, and they are not consequences of the hypothesis.
One more caveat, from experts on the October result: much of what a full proof would buy may already be available. "In a sense, the vast majority of consequences of RH already follow from quasi-RH," Radziwiłł told Quanta. If OpenAI’s quasi-Riemann proof holds up, the practical mathematical gain from the full hypothesis may be smaller than its fame suggests, and the significance of a proof would rest as much on its methods as on its conclusion.
The prize, the credit and the rules
The Clay Mathematics Institute’s $1 million prize is not won by posting code. Under the institute’s rules, revised in 2018, a proposed solution must first appear in a "Qualifying Outlet," essentially a recognized refereed journal or similar venue; at least two years must pass after publication; and the solution must have achieved general acceptance in the global mathematics community. Only then does the institute’s scientific advisory board consider it, usually with a special committee of experts. The institute does not accept direct submissions. By those rules, Friday’s Lean archive has not started the clock.
Only one Millennium Problem has been officially solved. Grigori Perelman posted his proof of the Poincaré conjecture on the arXiv in 2002 and 2003, never submitted it to a journal, and was nevertheless awarded the prize in March 2010, after experts spent years checking it. He declined the money, as he had earlier declined a Fields Medal. The Perelman precedent shows that the institute can be flexible about the form of publication if the mathematics is clearly right. It also shows how long the process takes.
Credit is a harder question than money. The note gives Claude the starring role in the final steps, but prizes and journal bylines go to people. When OpenAI announced its Navier–Stokes result in September, it said it did not intend to claim the Clay money, and the institute said only that the problem had "apparently been settled." How Clay would treat a proof largely found by an AI system, with a human guide who openly says he has not yet mastered all of it, is a question no rule book has had to answer before. The Fields medalists’ declaration in September was, in part, an argument about exactly this.
There is also a corporate dimension. Alpöge works at Anthropic, the proof was found with Anthropic’s model and computing resources, and the note says he consulted the company’s leadership before posting. Anthropic had not said, as of Friday morning, whether it considers the result its own, whether it will publish its own account as it did in August, or whether it will claim any prize.
What it would mean for mathematicians, and for AI
Set aside, for a moment, whether this particular proof is right. The way it was presented, a few pages of prose and 17 million lines of machine-checkable code, says something about where mathematics is heading, and the profession has been arguing about it all year.
The prediction that came true. In a 2023 essay, Terence Tao predicted that "2026-level AI, when used properly, will be a trustworthy co-author in mathematical research." He also warned that the profession was "largely unprepared" for what that would mean for journals, students and credit. Friday’s note reads like a test case for both halves of that forecast. It describes a human with a decades-old plan and a machine that "flew, and danced, and sparkled near that finish line," and it leaves every institutional question, from authorship to refereeing, unanswered.
Understanding versus certainty. Mathematicians have always valued proofs for the understanding they bring, not only for the truth they certify. The note itself acknowledges the gap: the author says he presents "a complete argument before having mastered a massive neighbourhood around its mathematics." If the code checks out, mathematics will for the first time hold a certified answer to its most famous question without anyone being able to say why it is true. The Harvard mathematician Melanie Matchett Wood warned earlier this year that "it will be easier for AI to convince humans it has a proof than to come up with a correct mathematical argument." Lean answers the first half of her worry. It does nothing for the second, the human need to understand.
The volume problem. OpenAI released 722 papers in one day in October. Friday’s note mentions 250,000 pages of drafts. No referee system was designed for that. The new Simons Foundation effort and OpenAI’s advisory group exist because mathematicians fear their field could be swamped by verified but unread results. The Fields medalists’ declaration in September made the same point more sharply, arguing that treating famous problems as AI benchmarks clashes with the purpose of mathematics.
The race. For the AI industry, famous mathematics problems have become a public scoreboard. OpenAI has Navier–Stokes, with an asterisk, and the quasi-Riemann Hypothesis. Anthropic had the two-thirds result in August and, if Friday’s claim holds, the biggest prize of all. The note’s unusual praise for OpenAI’s "good taste," and its comment that the Navier–Stokes work made the author realize that problem "was going to fall," hint at how closely the two labs’ mathematicians watch each other. The credit dispute between Buckmaster’s team and OpenAI over Navier–Stokes shows how quickly that rivalry can turn sour.
The upside. None of this is only a threat. If an AI system can produce a correct proof of the Riemann Hypothesis that humans can then study, mathematicians gain a new theorem and a new set of techniques at once, and the note argues those techniques reach far beyond this one problem. After the October result, Radziwiłł said the general feeling among number theorists was "one of relief," and Quanta noted that it would establish scores of results across number theory, geometry and theoretical computer science in a single stroke. A full proof of GRH would do far more. The author’s own hope, stated in the note, is that "those who love the mathematics will all be able to enjoy such transcendence in the future."
Reasons for caution
Friday’s claim deserves to be taken seriously. It also deserves skepticism, for several reasons that have nothing to do with anyone’s good faith.

Screenshot: X / @littmath
The history of failed claims is long. De Branges in 2004 and Atiyah in 2018 are only the most prominent. Serious mathematicians, not cranks, have announced proofs that collapsed under examination. The difference this time is the formal proof, but that shifts rather than removes the risk.
Formal proofs can prove the wrong thing. Earlier this year, Daniel Litt of the University of Toronto described submissions to an AI mathematics challenge that included "an enormous amount of garbage," including incorrect solutions that claimed to be formalized in Lean. Usually the problem was a mismatched statement or a hidden assumption. Checking that this archive proves Mathlib’s RiemannHypothesis, and nothing weaker, is the whole game.
The build had not finished. On Friday morning, nobody outside the author’s circle had publicly reported a complete, clean build of the project with an axiom check. Until that happens, the claim is a claim about code, not a verified proof.
No human-readable proof exists yet. The note itself says the author has not "mastered a massive neighbourhood around its mathematics." For a result this important, mathematicians will want to understand the argument, not just accept it. And a proof that no one can explain will be hard to build on.
The timing invites suspicion. The note was posted the day after Anthropic’s stock began trading. The note anticipates the objection and calls the "obvious cynical view" false. The only real answer is the verification itself, which does not depend on anyone’s motives.
A second, unverifiable claim came the same day. The Navier–Stokes paper with Buckmaster relies on a computer certificate that had not been released, so it could not be checked either. Two Millennium Problems on one Friday is extraordinary, and extraordinary claims need extraordinary evidence.
At the same time, there are real reasons this claim is different from the usual crank proof. The author is a well-regarded professional number theorist. Earlier AI results on the same problem, from both Anthropic and OpenAI, were checked and largely accepted by experts. The note names a specific technical idea in the specialists’ own language. And the code is public, so the claim can be falsified quickly. None of that makes it true. It does make it worth checking.
What we know, and what’s still unclear
What we know:
- A note posted on Levent Alpöge’s website on Nov. 20, 2026, claims a proof of the Riemann Hypothesis and the Generalized Riemann Hypothesis, found with Anthropic’s Claude.
- The proof was released only as Lean code, in 14 archive parts; an independent mirror counts 175,187 Lean files and about 17.4 million lines.
- A text scan of the archive found no "sorry," no new axioms and none of several common escape hatches.
- The main theorem is stated as rhOfAll : RiemannHypothesis, built against a pinned version of Mathlib.
- The note says the informal drafts exceeded 250,000 pages and that a readable write-up will follow.
- Earlier in 2026, AI systems from Anthropic and OpenAI produced expert-reviewed partial results on the same problem: two-thirds of zeros on the critical line, and a quasi-Riemann Hypothesis.
What’s still unclear:
- Whether the full project compiles, and how long a build takes.
- What #print axioms rhOfAll reports.
- Whether the archive’s definitions of the zeta function and the hypothesis match Mathlib’s exactly.
- Whether the 62 metaprogramming calls flagged by the mirror do anything unusual.
- What Anthropic says, and whether it will publish its own account.
- Whether the human mathematicians thanked in the note have checked any of the argument.
- How the Clay Mathematics Institute will treat a Lean-only, AI-assisted proof.
A short glossary
This story uses a lot of specialist language. Here is a reference list, in roughly the order the terms appear.
- Prime number: a whole number greater than 1 whose only divisors are 1 and itself, such as 2, 3, 5, 7 and 11.
- π(x): the number of primes less than or equal to x. For example, π(100) is 25.
- Li(x), the logarithmic integral: Gauss’s smooth estimate of π(x), built by adding up the "probability" 1/ln t of each number t being prime.
- Riemann zeta function, ζ(s): the function defined by the sum 1 + 1/2ˢ + 1/3ˢ + …, extended to the whole complex plane by Riemann in 1859.
- Nontrivial zeros: the zeros of zeta inside the critical strip. Their positions control the errors in prime counting.
- Critical strip and critical line: the strip of complex numbers with real part between 0 and 1, and the vertical line down its middle where the real part equals 1/2.
- Riemann Hypothesis: the statement that every nontrivial zero lies on the critical line.
- Dirichlet L-function and GRH: cousins of zeta that track primes in arithmetic progressions. The Generalized Riemann Hypothesis says their nontrivial zeros all lie on the critical line too.
- Quasi-Riemann Hypothesis: the weaker claim that all nontrivial zeros lie in some strip narrower than the critical strip. OpenAI claimed a proof in October 2026.
- Finite field and function field: number systems with finitely many elements, like arithmetic on a clock, and the curves built from them. The Riemann Hypothesis for function fields was proved by Weil and Deligne.
- Multiplicity: how many times a polynomial vanishes at a point, measured by how many of its derivatives also vanish there. Central to Stepanov’s proof.
- Automorphic form: a highly symmetric function attached to a group. The modern theory of these functions underlies much of number theory, including the proof of Fermat’s Last Theorem.
- Theta lift: a way of transporting automorphic forms from one group to another. The note’s key idea is a long chain of them.
- Lean: a programming language and proof assistant in which mathematical proofs can be checked by computer.
- Mathlib: Lean’s community library of formalized mathematics, which contains the official formal definition of the Riemann Hypothesis.
- Kernel: the small core program inside Lean that actually checks proofs.
- sorry and axioms: "sorry" is Lean’s placeholder for a missing step; an axiom is a statement assumed without proof. A trustworthy proof uses no sorries and only Lean’s three standard axioms.
- Millennium Prize Problems: seven problems named by the Clay Mathematics Institute in 2000, each with a $1 million prize. Only the Poincaré conjecture has officially been solved.
Questions and answers
Has the Riemann Hypothesis been proved?
Not in any sense the mathematical community recognizes yet. A proof has been claimed and released as computer code. If that code compiles cleanly, its axioms are standard and its statement matches the official definition, the hypothesis will have a formal proof, but experts will still need to understand it before treating it as settled.
Who is Levent Alpöge?
Alpöge is a 34-year-old number theorist. He studied at Harvard, graduating in 2014, spent a year at the University of Cambridge as a Churchill Scholar, and earned his doctorate at Princeton in 2020 under Manjul Bhargava with a thesis on rational points on curves. He was a Junior Fellow of the Harvard Society of Fellows, won the Morgan Prize for undergraduate research in 2015, and co-wrote a 2025 paper on Hilbert’s tenth problem. He later joined Anthropic, where he worked on the company’s August result on zeta zeros and its AI-assisted fluid-dynamics work with Buckmaster.
What did Claude actually do?
According to the note, the author set the overall strategy, an analogy with Stepanov’s function-field proof, and Claude supplied most of the technical machinery, especially in automorphic forms. The note says "Claude far outpaced me near the finish line" and that Claude and the author together "formalized the arguments without assumptions." The note does not say which Claude model was used or how much computing went into it. Anthropic’s August post about the earlier zeta result described an unreleased research model that ran dozens of subagents and millions of tokens.
Would a proof affect encryption or cryptocurrencies?
No. Encryption depends on how hard it is to factor particular large numbers, and a proof of the hypothesis provides no new way to do that. The myth comes from the fact that both topics involve prime numbers.
Can anyone check the code?
Yes, in principle. The archive and its mirror are public. In practice, compiling 17.4 million lines of Lean takes substantial computing power and expertise, and the meaningful checks, the axioms and the statement, need people who know Lean well. Expect the first credible verdicts from Lean specialists, such as the Mathlib maintainers and researchers like Buzzard, rather than from casual downloads.
Is this the first time AI has solved a famous open problem?
No. In 2026 alone, AI systems were credited with disproving Erdős’s unit distance conjecture, finding a counterexample to the Jacobian conjecture and, if it holds, settling the Clay version of the Navier–Stokes problem. But the Riemann Hypothesis is in a category of its own. For many mathematicians it is the problem, and a verified proof would be the clearest sign yet that AI systems can do research mathematics at the highest level.
Why does the "generalized" version matter so much?
Because most applications need it. The ordinary hypothesis controls primes overall, but many problems ask about primes of a special form, such as primes that leave a remainder of 1 when divided by 4, or primes in other arithmetic progressions. Those questions are governed by Dirichlet L-functions, and only the generalized hypothesis controls their zeros. The primality-testing, Goldbach and primitive-root results described above all depend on GRH, not just on the original statement. That is why the note’s claim to cover GRH would, if true, matter even more than the headline.
Why is it still called a hypothesis?
Riemann himself never named it. Later mathematicians called it a hypothesis, a statement assumed in order to see what follows from it, and the name stuck. Mathematicians usually call unproved statements conjectures, and a proved one becomes a theorem. If Friday’s proof is confirmed, the name will almost certainly survive anyway, the way Fermat’s Last Theorem was called a theorem for 358 years before anyone proved it.
How long until we know?
Two different clocks are running. The formal one could be fast: if the build succeeds and the axiom check is clean, Lean experts could give a preliminary verdict within days or weeks. The human one is slow: understanding a proof that touches this much of number theory could take years, which is about how long it took the community to absorb Perelman’s work. The Clay prize clock, two years after publication in a journal, has not started.
What happens next
Here is what to watch for in the coming days and weeks.
A clean build and an axiom report. The single most important next fact is whether someone independent of Anthropic publicly reports compiling the whole project and running #print axioms rhOfAll with only the three standard axioms in the output. On Friday the Hugging Face mirror listed such a build as in progress.
A statement audit. Lean experts will compare the archive’s definitions against Mathlib, line by line, and inspect every custom tactic and metaprogram.
An independent kernel. Replaying the exported proof in a second checker, as outside researchers did with OpenAI’s quasi-Riemann proof, would add another layer of confidence.
Word from Anthropic. The company published detailed accounts of its earlier mathematical results. Whether it does so here, and what it says about the role of its model, will matter both scientifically and, now that it is a public company, commercially.
A human write-up. The note promises one. Until it exists, the argument cannot be absorbed into mathematics the usual way, through seminars, journals and the slow work of other mathematicians building on it.
Expert reactions. Watch for statements from the people the note thanks, including Bhargava, Sarnak and Radziwiłł, and from the Fields medalists who signed September’s declaration.
The Navier–Stokes certificate. If Alpöge and Buckmaster release the certificate for their unforced blow-up claim, that result can be checked too.
The bottom line
For 167 years, the Riemann Hypothesis has been mathematics’ most famous unfinished sentence: a one-line guess about where a function equals zero that turns out to govern how evenly the prime numbers are spread. Generations of mathematicians proved real theorems about it, computers checked trillions of cases and nobody could finish it. On Friday, a number theorist at Anthropic said he and Claude had, and posted a 17-million-line Lean proof in place of a paper. Whether that is the most important moment in the history of mathematics or the most elaborate false alarm will be decided not by reputations or press releases but by a computer build, an axiom check and, eventually, human understanding. Either way, the year that began with AI solving olympiad problems is ending with AI at the doorstep of the hardest problem in mathematics.
Further reading

Prime Obsession by John Derbyshire
A readable history of the Riemann Hypothesis that alternates chapters of math with chapters on the people, from Euler and Riemann to the random-matrix connection.

Riemann’s Zeta Function by H. M. Edwards
The classic guided tour of Riemann’s 1859 paper, with an English translation of the paper itself; best for readers comfortable with calculus.

Amazon Echo Spot
A small smart display and alarm clock for a desk or nightstand, useful for timers and quick answers during long reading sessions.

Amazon Echo Dot
A compact smart speaker for reminders, timers and news briefings, including updates as stories like this one develop.

Apple AirTag 4-pack
Item trackers for keys, bags and laptop sleeves, for anyone who carries notebooks and devices between office, library and home.
