Key takeaways
- Levent Alpöge believes he and Claude have proved the Riemann hypothesis, a 167-year-old conjecture about how prime numbers are distributed.
- Two independent proof checkers accepted the formal argument, while Claude’s attempt to explain it in writing ran to more than 250,000 pages.
- About three months separated Claude’s reported advance on one part of the Riemann problem from Alpöge’s claim to have solved the whole thing.
- For an investor assessing an AI scientist, one revealing question is which experiment a researcher chose to run because of its prediction.
Claude may have helped solve a problem that has eluded mathematicians for 167 years. On Friday, Levent Alpöge, a mathematician at Anthropic, announced that he and Claude had found what he believes is a proof of the Riemann hypothesis. There are good reasons to take the claim seriously. And if it holds up, mathematicians will face an unusual situation. A computer can already check an argument that its own author is still learning to explain.
My favourite detail is buried in a footnote to Alpöge’s four-page announcement. He tried having Claude write the argument out and got more than 250,000 pages. Curious what such a book might read like, he looked at the first theorem. Claude began by carefully explaining that if a real number differs from one-half, it must be either greater than one-half or less than one-half. You can see how it got to 250,000 pages.
Alpöge spared his colleagues the manuscript. What he released was roughly 17 million lines of code in Lean, a language designed for checking mathematical proofs. That decision gave other mathematicians something they could test immediately, and it gets to what I find most interesting about this story. How quickly can we gain confidence in a discovery, and how long does it take to understand what we’ve discovered?

What was Riemann trying to prove?
Pick a number, say a million, and ask how many prime numbers there are below it. You could count them. You could also use a formula that gives a good estimate, and the estimate works surprisingly well as you move out to a billion, a trillion and beyond. The interesting question is how wrong that estimate can be. Could the primes have a stretch so unusually sparse or dense that the count wanders much further from the prediction than anyone expects?
In 1859, Bernhard Riemann connected that question to a seemingly distant piece of mathematics. A function called the zeta function takes the value zero at certain points. Riemann proposed that all its non-trivial zeros lie on one particular straight line. If they do, mathematicians get remarkably tight control over the error in their prime counts. The first ten trillion of those zeros have been checked and found on the line. The difficulty is the word all. Even ten trillion examples leave infinitely many to go. The Clay Mathematics Institute’s explanation is a good introduction to why this became one of its million-dollar Millennium Prize Problems.
So where does the claim stand? As of 21 November, MIT mathematician Andrew Sutherland reports that two independently written proof checkers accepted the full formal argument. A separate comparison also confirmed that the five theorem statements in the submission match the mathematical challenge they were supposed to answer. Some checks were still running, and Clay continued to list the problem as unsolved. That is a much more interesting state of affairs than the usual weekend announcement that someone on the internet has cracked Riemann. Other people can run the checks, inspect the assumptions and report exactly what passed.
From Olympiad medals to open problems
Two summers ago, the headline was that AI could compete with exceptionally talented teenagers. Then the questions changed.
In July 2024, Google DeepMind’s AlphaProof and AlphaGeometry 2 solved four of the six problems at the International Mathematical Olympiad, a silver-medal performance. Experts first translated the questions into a form the systems could work with. A year later, an advanced version of Gemini Deep Think reached the gold-medal standard, solving five of six from the original natural-language questions within the competition time limit. For anyone tracking model progress, that last detail matters as much as the medal. A layer of human assistance had fallen away.
In May 2026, an OpenAI research model found a counterexample to a conjecture Paul Erd?s had posed in 1946. Imagine scattering dots on a sheet of paper and counting how many pairs sit exactly one unit apart. Erd?s had conjectured a limit on how fast that count could grow as you added dots. The model found a construction that broke the proposed limit. An Olympiad committee knows its problems have solutions before the contestants sit down. In this case, mathematicians had spent 80 years trying to find out whether the proposed answer was even true. Quanta tells the story of the Erd?s result here. Quanta Magazine
Then, in August, came a result that makes this week’s announcement easier to put in perspective. Anthropic reported that a research version of Claude had raised the proven lower bound on the fraction of zeta zeros lying on Riemann’s line, from 41.6% to 67.2%. This was a guarantee about the zeros as a whole, which is why it mattered even though trillions of individual zeros had already been checked. Roughly three months later, Alpöge announced a proposed proof covering all of them.
Read those dates together. A year separated the silver and gold Olympiad performances. Ten months later came the Erd?s counterexample. Within another six months we had both the stronger Riemann bound and a claim to the whole thing. I’d hesitate to plot an exponential curve through a handful of very different results. What has changed is the conversation you can have with a researcher. In 2024, you could reasonably ask whether a model could follow an argument that the best students could find. Today you can ask an expert which move the model found that surprised him.
What changes for science?
Alpöge gives a glimpse of what that collaboration actually looked like. As a student, he imagined taking on Riemann decades later and deliberately learnt techniques he thought might one day come in handy. Working with Claude, he recognised that two ideas from a simpler version of the problem might fit together and steered the work towards a setting where they could try them. Later, Claude suggested a specific mathematical choice that astonished him. I find that account much more revealing than trying to assign a percentage of the credit to the human or the machine. Years of acquired judgement shaped the question the model got to answer. alpo.ge
How much of this travels beyond mathematics? Consider the AlphaFold Protein Structure Database, which already makes more than 200 million predicted protein structures freely available. A biologist can start a Monday morning with a plausible shape for a protein that might once have taken a considerable effort to characterise. Whether it behaves as expected in a living cell is still an experimental question. Mathematics has Lean. A biologist may have to book time on an instrument, order a reagent and wait for cells to grow. Better models can make that experiment more informative even when they can’t make it finish any sooner.
As an early-stage investor, that changes the question I’d ask a founder building an AI scientist. Show me a prediction your system made, the experiment someone ran because of it, and what the system learnt when the result came back. If a lab can afford to run only three experiments this month, helping it choose the right three could be worth far more than generating another thousand plausible ideas.
There’s one more discovery I’d love to see come out of the Riemann work. Alpöge suspects the argument can be simplified substantially. Suppose someone gets those 250,000 pages down to ten, and on page seven there’s an idea a graduate student can take into a completely different problem. A proof would settle a question from 1859. That page might tell us what to ask next.
- Did Claude help prove the Riemann hypothesis? - November 21, 2026
- Weekly Firgun Newsletter – November 20 2026 - November 20, 2026
- Why AI Labs Are Buying the Data Behind Real Work - November 18, 2026
Sources
- Levent Alpöge’s original announcement
- MIT, Andrew Sutherland’s independent verification results
- Clay Mathematics Institute, the Riemann hypothesis explained
- Anthropic, Claude’s earlier result on the Riemann zeta function
- Google DeepMind, Gemini reaches gold-medal standard at the Maths Olympiad
- Quanta Magazine, why the Erd?s problems are falling to AI

