How the Riemann Hypothesis Proof Was Certified
On Friday morning, a mathematician posted a four-page note on his personal website. By Saturday, the most respected proof-checking community in mathematics had certified his Riemann hypothesis proof.
Levent Alpöge, a mathematician at Anthropic, had announced that he and the company’s AI model Claude had proved the Riemann hypothesis. It was the boldest claim in mathematics in a century.
The next day, the Lean and Mathlib communities published a joint statement. Their words were careful, but their meaning was not. “We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct,” they wrote. “We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct.”
It was a moment no one in mathematics had expected to see this decade. And it was only half the story.
Hours after Alpöge’s note appeared on Friday, a second proof of the same problem landed online. This one was written entirely by a human, over twelve years, with no AI involved at all.
What Happened on Friday
Alpöge’s note was short. In a single sentence, he wrote that he believed he had, working with Claude, discovered a proof of the Riemann hypothesis. He added that the same method settled the Generalized Riemann Hypothesis, a broader version of the problem. We covered the original announcement when it landed on Friday.
There was no conventional paper. Instead, Alpöge released the argument as computer code written in Lean, a proof assistant that checks every logical step. The archive was split into fourteen parts and ran to roughly seventeen million lines.
Written out in ordinary mathematical prose, the argument would stretch to more than 250,000 pages. That is why the code, not a paper, was the proof.
Alpöge was candid about what that meant. He said plainly that he did not yet command all of the mathematics inside it. Claude, he explained, had pulled far ahead of him in the final stretch.
The work had built up over weeks. Alpöge had been working with Claude across a range of mathematical areas since October 6. The decisive stretch lasted forty-one days, ending on November 16.
It was not his first headline with Claude. Earlier, he had used the model to produce a counterexample to the Jacobian conjecture, an eighty-seven-year-old problem in algebraic geometry.
How the Independent Checks Worked
A claim this large invites suspicion, and the first checks arrived within hours. Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, downloaded the code and compiled it on his own machine.
It compiled without errors. More importantly, the Lean kernel rechecked the proof and found that it rested only on the three standard axioms of mathematics. Sutherland’s verdict was blunt: the chance that anything was fishy, he said, was approximately zero.
But compiling is only the first question. A subtler worry is whether the code proves the Riemann hypothesis itself, rather than a slightly different statement that happens to look similar.
That is where the weekend’s certification came in. The Lean and Mathlib communities wrote out the hypothesis in a separate file, using only Mathlib, the shared library of mathematics. Then they compared it, line by line, with the theorem Alpöge had proved.
Two independent tools confirmed the match. The statement in the code really is the Riemann hypothesis as mathematicians understand it.
Sutherland’s verification page, updated through the weekend, tells the rest of the story. All 175,187 modules in the proof compile with zero errors. The five headline theorems depend only on the three standard axioms.
Two independently written kernels, called nanoda and con-leche, accepted the full exported proof. An independent rebuild in a sandboxed machine, with no network access, reproduced the export byte for byte.
Some final checks were still running at last report. The Lean kernel’s own replay of the export, and a full comparator run, had not finished. But the checks that matter most had already passed.
The Second Proof: Twelve Years, No AI
While number theorists were still downloading Alpöge’s files, a second proof appeared. Ravi Vakil, a mathematician at Stanford University and president of the American Mathematical Society, posted it on his own blog.
Vakil wrote the post in a rush, he said, from an AMS board meeting, minutes before chairing a session that would last hours. The manuscript came from Alessandro Maria Masullo, a postdoc at Stanford who received his PhD there in 2020.
Masullo’s draft runs to 940 pages. It claims the Generalized Riemann Hypothesis, the same broader statement Alpöge says he has settled. Vakil’s note carried a striking line: “It is all human, no AI.”
The two efforts could hardly have more different origins. Masullo’s work began twelve years ago, when he was a graduate student under Brian Conrad at Stanford. Conrad, one of the most respected number theorists alive, had spent months checking the draft and suggesting revisions.
The plan had been to spend a year polishing the writing, then announce it after running it past enough people. Alpöge’s announcement changed that. Conrad and Masullo decided to release the current draft instead.
Masullo has already acknowledged mistakes in his manuscript. His proof, unlike Alpöge’s, has not been machine-checked. It now faces the slow, human process of peer review.
Why the Riemann Hypothesis Proof Matters
The Riemann hypothesis was posed in 1859 by the German mathematician Bernhard Riemann. It concerns a hidden order in the distribution of prime numbers, the numbers divisible only by one and themselves.
The hypothesis predicts exactly where a mathematical function called the zeta function equals zero. If it is true, the primes are distributed in a far more orderly way than they first appear.
Thousands of mathematical results depend on it. If the hypothesis falls, a vast web of theorems becomes solid ground instead of conditional truth. That is why it is one of the seven Millennium Prize Problems chosen by the Clay Mathematics Institute, each carrying a one-million-dollar prize.
Riemann himself was born two hundred years ago this September. The confirmation of either proof would be a spectacular bicentenary present, as one report put it.
The year had already been extraordinary for AI in mathematics. Earlier in 2026, mathematicians using AI tools announced a solution to another Millennium problem, related to the Navier-Stokes equation.
In August, Claude raised a key bound on the zeros of the zeta function from 41.6 percent to 67.2 percent. In September and October, OpenAI researchers published a quasi-proof of the Riemann hypothesis itself.
But a full, machine-checked proof of the real thing would stand above all of it.
What Happens Next for the $1 Million Prize
Certification is not the same as acceptance, and the Clay Mathematics Institute still lists the problem as unsolved. Under its rules, a prize needs publication and expert acceptance, a process that takes at least two years.
That work will take time. The argument is, in the words of one leading mathematician, crazily complicated. It may take years before anyone can explain in ordinary language why it works.
The Lean communities have asked mathematicians to help explain the new ideas inside it. No human has yet read and understood the proof.
James Maynard, a Fields medallist at Oxford, is optimistic. “I suspect that the proof will turn out to be correct,” he said, even though it now exists only as incomprehensible code.
Masullo’s draft faces a different road. Human reviewers must now read 940 pages, find the errors he has already acknowledged, and decide whether the core argument survives them. That process has barely begun.
Either way, the mathematics community now has two serious, independent claims to the same prize. That has never happened before in the history of the problem.
The Doubts That Remain
Not everyone is celebrating without reservation. Andrew Granville, a number theorist at the University of Montreal, called Alpöge’s four-page note garbled. He said Alpöge should have prepared a clearer write-up before making the announcement.
Granville also questioned the timing. The announcement came days after Anthropic’s stock market debut, when the company’s valuation had dropped sharply. Alpöge has denied any cynical motive in the timing.
The timing question is real, but it does not touch the mathematics. A proof that compiles, rests on only the standard axioms, and matches the standard statement is either right or wrong. The machines have now said it is right.
It is the same kind of machine-checked verification that has already been used to prove real software secure.
What the machines cannot do is explain it. That is the work ahead, for humans, with or without AI. Two proofs arrived in a single day. Only one has been certified so far, and neither has yet been understood.
The next chapter of this story will be written slowly, by mathematicians reading, arguing, and, eventually, agreeing.
According to Scientific American, independent researchers have now verified that the proof does what Alpöge claims. The magazine’s headline calls the proof almost certainly true.



