Anthropic Mathematician Says Claude Helped Prove the Riemann Hypothesis — What Is Actually Verified

On November 20, 2026, a mathematician at Anthropic said he believes he has, working with Claude, found a proof of the Riemann Hypothesis. The claim landed in a four-page note posted on his personal website, and it concerns one of the most famous unsolved problems in mathematics: a question about the distribution of prime numbers first posed by Bernhard Riemann in 1859, and one of the Clay Mathematics Institute’s Millennium Prize Problems, carrying a $1 million prize.
That is a very large claim, and the first thing to say about it is that it is a claim, not an accepted result. The proof has been checked by machines in important ways, but no human has read the whole thing, no peer review has happened, and the Clay Mathematics Institute still lists the problem as unsolved. This article sets out what was claimed, what has actually been verified, and what would still have to happen before anyone could call it a theorem.
What Was Actually Claimed
Levent Alpöge, a mathematician at Anthropic, wrote in his note: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” He says the same method also proves the Generalized Riemann Hypothesis, the wider version of the problem. He consulted Anthropic CEO Dario Amodei before announcing, and Anthropic itself has issued no formal statement.
There is no conventional paper. The proof exists only as Lean code, running to about 17 million lines and shared as a 14-part archive. A written version, Alpöge says, would run past 250,000 pages. That scale is the heart of the story. The argument is so long that it cannot be read by a person in any realistic sense, and Alpöge is open about that: he says he does not yet command all the mathematics in it, and that Claude pulled ahead of him in the final stretch.
The decisive work ran 41 days, from October 6 to November 16, 2026. That is a short window for a problem that has resisted the best mathematicians for more than a century and a half, and it is one reason the claim drew such fast attention — and such fast scrutiny.
What Has Been Checked So Far
The verification so far has been done by machines, and it moved quickly. MIT mathematician Andrew Sutherland and others rebuilt the Lean code from the archive, and by November 21 the Lean and Mathlib communities reported that separate checkers — nanoda, con-leche, and lean4lean — had confirmed that both the statement of the theorem and the proof match Mathlib’s standard definition of the Riemann Hypothesis.
That is a serious check, and it is worth understanding what it does and does not mean. The checkers confirmed that the code proves the statement as Mathlib defines it. They did not, and could not, tell us whether the argument makes sense to a human reader, or whether the definitions line up with the problem as mathematicians understand it outside Lean. Machine checking answers a precise question: does this code, run through these tools, establish this formal statement? On that question, the early answer is yes.
Some final end-to-end comparator and kernel replay checks were still running on November 21, so even the machine verification was not fully complete at the time of writing. Those checks are the last step in confirming that the code does what it claims from end to end, and their results matter before anyone draws a firm line under the verification.
Why This Is Not Yet a Theorem
A proof becomes a theorem when the mathematical community accepts it, and that has not happened here. No human has read the full proof, and at 17 million lines of code, no human is going to. There has been no peer review, and the Clay Mathematics Institute still lists the Riemann Hypothesis as unsolved.
The prize rules make the gap clear. The $1 million Millennium Prize needs journal publication, two years of scrutiny, and general acceptance by the mathematical community. None of those steps has happened. A machine-checked formal proof is a strong piece of evidence, but it is not the same as a result the community has examined, understood, and accepted.
That gap is not a technicality. Mathematics runs on understanding as well as correctness. A proof that no one can read leaves open questions that a checker cannot answer: whether the formal statement captures the intended problem, whether the argument contains steps that are formally valid but mathematically empty, and whether the whole thing could be rebuilt in a form a person can follow. Until those questions have answers, the honest description is a serious machine-checked claim, not a proven theorem.
A Rival Proof Appeared the Same Day
Hours after Alpöge’s announcement, a second claim entered the picture. Ravi Vakil of Stanford University, the president of the American Mathematical Society, shared a rival human-authored draft proof by Alessandro Masullo. The timing was striking: two claimed proofs of the same 166-year-old problem surfacing within hours of each other, one built with an AI system over 41 days, the other written by a human author in the usual way.
The rival draft does not settle anything either. It is a draft, and it faces the same road any proof faces: reading, checking, peer review, and acceptance. But its arrival shows how crowded and fast-moving this moment is. The Riemann Hypothesis has been unsolved since 1859, and now two claimed proofs are being talked about on the same day.
The Earlier Result That Set This Up
This claim did not come out of nowhere. In August 2026, a Claude research version raised the guaranteed share of zeta zeros on the critical line from 41.6% to 67.2%. That was a real, measurable advance on a quantity tied to the Riemann Hypothesis, and it showed that AI systems were already making headway on the mathematics around the problem.
Going from a stronger partial result to a claimed full proof in a few months is a big jump, and it is the jump that the mathematical community is now trying to assess. The August result was the kind of progress mathematicians know how to judge. The November claim is a different kind of object altogether: a proof too long for any person to read, built in weeks, and verified so far only by other machines.
What to Watch Next
Three things will decide how this story ends. First, the remaining end-to-end checks on the Lean code need to finish, so the machine verification is complete rather than nearly complete. Second, mathematicians need time to study what the code actually does, and whether the formal statement it proves is the problem as it is understood. Third, the rival draft by Masullo will go through its own scrutiny, and the two claims will be weighed against each other.
For now, the careful way to describe it is this: a mathematician working with Claude has produced a machine-checked formal argument that, if it holds up, would settle the Riemann Hypothesis and its generalised form. That is a remarkable thing to be able to say at all. It is not the same as saying the problem is solved. For the wider reporting, see the Phys.org report on the Claude Riemann claim and the ExplainX breakdown of what is claimed versus what is verified. And if you want a sense of how AI models are being used for hard technical work closer to home, our report on Inkling 2 Mini, a local AI model and our piece on Google Gemini 4 Neon both show the same tools being put to very different jobs.
The next few weeks will tell us whether this becomes a footnote or a turning point. Either way, it is already a test case for a question mathematics has never had to answer before: what do we do with a proof no human can read?
