Explainer · Artificial intelligence
The Riemann hypothesis, the million-dollar problem Claude says it solved
Levent Alpöge, a mathematician at Anthropic, says Claude has proved the Riemann hypothesis, the most famous open problem in mathematics. For now there are 14 parts of Lean code and no independent verification. Here is what the theorem says, why it matters and what is still missing.
On Friday 20 November, Levent Alpöge published a short note on his website: together with Claude, he said, he had proved the Riemann hypothesis. He did not present a paper, a plain-language version, or a PDF with a proof in the classical sense. He presented code: a Lean formalisation split into 14 parts which, by his own description, is equivalent to some 250,000 pages of conventional mathematics.
The news is enormous if the code does what it says. And that is the problem: nobody has yet checked it independently. Not the Clay Mathematics Institute, which still lists the problem as open, not a peer-reviewed journal, not an external team of mathematicians has confirmed that the formalisation compiles, that it proves exactly the hypothesis and that it does not rely on extra axioms. It is an announcement, not an accepted result.
Alpöge works at Anthropic, the company behind Claude. That does not invalidate anything, but it does force a separation between two things: what the code says and what the marketing says. There is no record of Anthropic presenting the work as a verified result of the company. The mathematical community has not validated it either. The honest sentence today is “there is a published claim”, not “the hypothesis has been proved”.
What “formalised” means. A Lean proof is code that a program checks step by step. That is a huge advantage over a PDF: the computer does not accept a logical leap. But it only counts if the code compiles, if the statement really is the Riemann hypothesis and if there are no tricks such as a sorry (Lean’s “I’ll skip this bit”) hidden somewhere.
What the hypothesis actually says
The Riemann zeta function is defined, for complex numbers with real part greater than 1, as an infinite sum: 1 + 1/2s + 1/3s + 1/4s and so on forever. In 1859 Riemann showed that the sum can be extended to almost the whole complex plane, and that its zeros — the values of s that make the function equal zero — come in two families.
The first is boring: the negative even integers (−2, −4, −6…), known as the trivial zeros. The second is the one that matters. These are the non-trivial zeros, and every one computed so far — more than ten trillion — has the same real part: ½. The hypothesis states that all non-trivial zeros, without exception, fall on that vertical line, the critical line. It does not say they sit anywhere in the plane: it says they all sit exactly on the same axis.
That sentence, short enough for a tweet, connects to the prime numbers. The zeta function is how number theory packs all the information about the primes into one object. If the zeros are where Riemann said, the error when counting primes up to a number x is as small as it can possibly be. In formula form: the number of primes up to x, written π(x), is approximated by the logarithmic integral Li(x) with an error of order √x times the logarithm of x. Without the hypothesis, the best we can guarantee is a much larger error; with it, the error drops to the square root.
The score behind the noise
If you look at the list of primes — 2, 3, 5, 7, 11, 13, 17, 19, 23, 29… — it looks like noise. The gaps change without warning: one, two, four, two, four, two, four, six. There is no visible rule saying when the next one appears. That is why primes have fascinated mathematicians for centuries: they are the purest example of something that looks random but is not.
The Riemann hypothesis is the claim that this noise follows a score. The zeros of the zeta function are the frequencies in the score; the primes are the notes you hear. The more orderly the frequencies, the more regular the melody. Riemann did not supply the proof, but he wrote down the instrument with which it could be heard. Almost every advance in number theory over the past 160 years has consisted of tuning that instrument.
The analogy has a limit, and it is worth saying: it is not that primes “sing” in some mystical sense. It is that the zeta function is a machine for counting primes, and its zeros are the gears. If one gear slips out of place, the count drifts. The hypothesis says no gear slips.
Why it matters (and why it will not break your bank)
The hypothesis is one of the Clay Institute’s seven Millennium Problems, with a one-million-dollar prize. But the prize is the least of it. The point is that hundreds of theorems today begin with the same caveat: “assuming the Riemann hypothesis”. They are conditional results: if the hypothesis is true, then such a bound holds, such an algorithm is fast, such an estimate is valid. A proof would turn all those conditionals into unconditional theorems.
And cryptography? Here we should slow down. The Riemann hypothesis does not break RSA on its own. RSA rests on the difficulty of factoring large numbers, and no theorem says that a proof of Riemann factors anything in polynomial time. What it does is sharpen the bounds of some algorithms and guarantee the behaviour of primality tests that today work under the generalised Riemann hypothesis. It is a gain in knowledge, not an internet blackout.
The real value is another one: understanding why primes, which show no apparent pattern, obey a global law. It is the same kind of surprise physics met when it discovered that the irregularities of the planets are explained by a law of gravitation. The hypothesis says the disorder of the primes is an illusion of scale.
How a proof gets checked in 2026
Traditionally, a proof is accepted when a group of specialists reads it, understands it and finds no error. With Lean, the process changes: the program’s kernel verifies every step, so the question is no longer “is the reasoning right?” but “does the code compile and is the statement the right one?”. That is an important difference, because it moves the mistrust from the logic to the translation.
The checklist is concrete. First: the repository compiles from scratch, with no private dependencies or manual steps. Second: no sorry is left and no axiom is added by hand; Lean can print the axioms each theorem relies on. Third: the formal statement is the Riemann hypothesis and not a weaker version, a special case or a reformulation that is equivalent only under assumptions. Fourth: independent mathematicians review both the code and the translation. Fifth: a readable version exists, in ordinary mathematical language, published in a peer-reviewed journal.
The last point is not bureaucratic fuss. Clay’s rules require the result to appear in a peer-reviewed publication and two years to pass before the prize is awarded. An enormous formalisation can be correct and still fail that test. And there is a precedent worth remembering: in 2018, Michael Atiyah presented an alleged proof in Heidelberg based on his Todd function. The room filled up, the story went around the world and the proof was never accepted: there were not enough details and the community found problems.
What to watch in the coming days
The first thing is the code. If the 14 parts can be cloned, compiled and verified with a public version of Lean and mathlib, the announcement moves up a notch. If it needs a special environment, a private binary or an unpublished branch, verification gets harder. The second is the statement: the community will ask to see the first line of the theorem, the one that says what is being proved. The third is Anthropic’s silence or confirmation: a company with a verified proof of Riemann would be expected to present it in more detail than a personal note.
The how matters too. If Claude has produced a complete formalisation, that is a milestone for automated proving, even if the proof turns out to have a flaw. Moving from “the model writes plausible text” to “the model produces code a verifier accepts” is a change of category. The interesting question is no longer only whether Riemann is true, but how much of a 250,000-page proof a machine can generate and check.
In the meantime, the reasonable position is the mathematical community’s: interest, scepticism and patience. A claim is not a theorem. And an unaudited formalisation is a claim written in a very precise language.
No map this time
This explainer has no map. The Riemann hypothesis has no geography: there are no countries, cities or points that add real context. Forcing a decorative map would be the opposite of what we try to do at Maña Maps. If you want to explore data with territory, the gallery gathers the published maps and each page explains its source and date.
This article describes a public, unverified announcement. It does not claim that the Riemann hypothesis has been proved. Quotes and figures come from the linked sources and from the note published by the author on 20 November 2026.
Sources
- Levent Alpöge (20 November 2026): personal note on the Lean formalisation. Description of the announcement, the 14 parts and the equivalence to some 250,000 pages.
- Clay Mathematics Institute: “Riemann Hypothesis”. Official statement, status of the problem and Millennium Prize rules.
- Lean Prover: official documentation and mathlib Community. How the verifier and the library of formalised mathematics work.
- Wikipedia: “Riemann hypothesis”. Definition of the zeta function, trivial and non-trivial zeros, and the link to prime counting.
- Scientific American and New Scientist. Coverage of the announcement and context of the community debate.
- Wikipedia: “Michael Atiyah”. Precedent of the alleged proof presented in Heidelberg in 2018.
- Maña Maps gallery. Published maps with data source and date.