Primes, cracked open16 minBuried in OpenAI's 700-paper math release was a proof that the Riemann zeta function has no zeros beyond 7/8, the quasi-Riemann hypothesis. Here's what that means in plain English, why experts are stunned, and what it does to the future of being a mathematician.Read now →

An Anthropic Mathematician Says Claude Helped Him Prove the Riemann Hypothesis in Six Weeks. The Written Proof Would Run 250,000 Pages
On November 20, Anthropic's Levent Alpöge posted a short note claiming that he and a non-public version of Claude had proved the Riemann Hypothesis, with the entire argument written in machine-checkable Lean code. Independent checkers are passing it so far. Here's what was claimed, what has actually been verified, why mathematicians are both stunned and uneasy, and what it could mean for AI building better AI.
For 167 years, the Riemann Hypothesis has been the problem mathematicians name when they want to describe the hardest thing in their field.
It carries a $1 million Millennium Prize. Thousands of published results quietly assume it is true. Some of the greatest minds of the last two centuries tried to prove it and failed.
On Friday, November 20, 2026, a mathematician at Anthropic posted a four-page note on his personal website that opened with a sentence few people expected to read in their lifetime:
"I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis."
He says the key work took about six weeks.
If the claim holds, it is the biggest result in the history of mathematics to involve an AI, and arguably one of the biggest results in mathematics, full stop. If it doesn't, it will be the most elaborate and most carefully checked false alarm the field has ever seen.
Right now it is somewhere in between, and that in-between state is the story. Here is what was claimed, what has actually been checked, why some experts are furious about how it was released, and why the implications reach far beyond prime numbers.
What was actually announced
The mathematician is Levent Alpöge, a member of Anthropic's technical staff. AP and the BBC both describe the announcement as a short note on his website, posted the day after Anthropic's stock market debut.
The note makes three escalating claims:
- The Riemann Hypothesis itself, for the Riemann zeta function.
- A generalized version (GRH), which he says the method extends to easily.
- An expectation that with more time the approach will reach the extended Riemann Hypothesis as well.
According to the BBC, the work was done with a non-public version of Claude, not the model you can chat with today.
The detail that separates this from the long history of crank proofs is the format. Alpöge did not post an essay asking the world to trust him. He released the argument as code in Lean, a programming language for mathematics in which every logical step has to be accepted by a small, strict checking program called a kernel. If one step is wrong, the proof simply does not compile.
The Lean release is enormous. AP reports it spans roughly 17 million lines of code. Alpöge says the natural-language version Claude produced would run to more than 250,000 pages, which is why he released only the formal version. He said it would be inappropriate to post a massive, unreadable document, and that he plans to write a proper human manuscript separately.
The note also has a timeline. By his account, the first key idea appeared on October 6, and the generalized result fell on November 16. That is roughly six weeks for a problem first posed in 1859.
The Riemann Hypothesis in plain English
You don't need the math to understand why people care.
Prime numbers, the numbers divisible only by 1 and themselves, look random: 2, 3, 5, 7, 11, 13, 17 and on forever, with no obvious rhythm. In 1859 Bernhard Riemann showed that the way primes are spread out is controlled by a single function, now called the zeta function, and specifically by the points where that function equals zero.
Riemann guessed that all the important zeros line up on one straight line. That guess is the Riemann Hypothesis.
Nobody could prove it. But mathematicians were so confident it was true that they built on top of it anyway. Huge stretches of number theory begin with the phrase "assuming the Riemann Hypothesis." A proof would turn all of those conditional results into settled ones at once.
That's why this is not "one more puzzle." It is a foundation stone that much of modern number theory has been standing on without being sure it was there.
Who is Levent Alpöge?
This is not an anonymous internet post. Alpöge is a trained number theorist who studied at Harvard and gave research talks at places like the Institute for Advanced Study before joining Anthropic.
People who know him describe him as one of the most gifted students of his Harvard class. By his own account, he had been planning an attack on this problem since his student days.
That matters for two reasons. First, it makes the claim far more credible than a random submission. Second, it shapes how to read the "AI did it" headline. This was not someone typing "solve the Riemann Hypothesis" into a chat box. It was a specialist who knew which tools mattered, steering a powerful model into the right setup.
Then, in his telling, the model took over.
"Claude flew and danced and sparkled"
The most quoted part of the note is how Alpöge describes the final stretch.
He writes that Claude far outpaced him near the finish line, and goes further: Claude "flew and danced and sparkled" there. He says he guided them both into the right framework and then Claude "began to leave me in awe."
He highlights one step he expects will make other mathematicians fall out of their chairs. At a key point the argument needed a chain of mathematical objects, and the question was how long that chain should be. Claude proposed the answer. His reaction, as he describes it, was essentially: of course, what else could it be?
That is the kind of move we normally call insight. It's the kind of thing prizes are given for.
He also admits something remarkable: the proof uses deep areas of mathematics he has not personally mastered. In his words, the moment has put him in the "rather extraordinary situation" of presenting a complete argument before mastering the neighborhood of mathematics around it.
What has actually been verified so far
Here's where you should be precise, because the internet wasn't.
Within a day, MIT mathematician Andrew Sutherland set up a public verification page. As of its update on the morning of November 21 (UTC), it reported:
- All 175,187 Lean modules compile with zero errors.
- The five headline theorems depend only on Lean's three standard axioms. No extra assumptions were smuggled in.
- The Mathlib library files used are byte-for-byte identical to the official ones.
- Every module passed a kernel replay.
- Two independently written kernels, nanoda (written in Rust) and con-leche, accept the full exported proof.
- The theorem proved matches Mathlib's own definition of the Riemann Hypothesis.
That is an extraordinary amount of checking for day one. No major proof claim in history has arrived with this much machine verification attached.
But the same page lists what is not finished:
- The official end-to-end Comparator run was still in progress.
- Lean's default and "paranoid" kernel checks on the exported proof had not returned a verdict.
- The build was done on machines without a sandbox, and a full comparison with a sandboxed build had not been done.
There is also an important scope note. Sutherland specifies that the "GRH" in the code covers Dirichlet L-functions only, not the generalized Riemann Hypothesis in its broadest sense. The headline claim about the original hypothesis is unaffected, but "a whole family of problems fell at once" is bigger than what has been checked.
So the honest status is: very strong machine evidence, not yet full acceptance. Nature's coverage makes the same distinction.
Why some mathematicians are angry anyway
Not everyone is celebrating.
Andrew Granville, a leading number theorist at the University of Montreal, told Nature that Alpöge should have prepared a clearer write-up before announcing, and called the released document "garbled." Nature also notes the timing: the announcement landed just days after Anthropic's IPO.
That criticism deserves to be taken seriously for three reasons:
- Correct is not the same as understood. A Lean proof can tell you a statement is true without telling any human why.
- Formal statements can be subtly off. The whole value of the check depends on the formal theorem meaning exactly what mathematicians mean. That is why the Comparator step matters.
- Commercial incentives are real. A company that just went public benefits enormously from a "Claude solved Riemann" headline. That doesn't make the proof wrong. It does mean independent scrutiny is non-negotiable.
None of these objections points to an error. They are about process, readability and trust. But they explain why the official verdict will take longer than the excitement.
Almost nobody saw this coming this fast
The speed is what has stunned forecasters.
The community forecast on Metaculus, the prediction site where thousands of people estimate future events, reportedly put a resolution of the Riemann Hypothesis around 2035 as of this week. Terence Tao, widely considered one of the greatest living mathematicians, has said AI might help crack problems like this before 2050.
Tao did see the direction early. He predicted that by 2026 AI would be a genuine collaborator in research-level mathematics, doing computations, verifications and some of the ideas. What almost nobody predicted was that it would arrive at the single most famous problem in the field this year.
Public bets tell the same story. Forecaster Greg Burnham had publicly bet that no Millennium Prize problem would fall within a year. After the announcement he posted that he had conceded and paid.
Look at the curve over just four years:
- 2023: language models stumbled on multi-digit arithmetic.
- 2024: they solved only a handful of high-school competition problems.
- 2025: AI systems reached gold-medal level at the International Mathematical Olympiad.
- Early 2026: models began cracking niche open problems.
- September 2026: OpenAI announced an AI-found solution to a forced version of the Navier-Stokes problem. We covered the follow-up physics result in AI just put the strong force on solid math.
- November 2026: a claimed proof of the Riemann Hypothesis.
Alpöge himself has a recent track record here. Earlier this year he used Claude in work on the Jacobian conjecture, a problem dating to 1939. The problems are getting bigger every few months. For a sense of the step just before this one, see our earlier piece on AI's quasi-Riemann result.
According to a friend, Alpöge once estimated the Riemann Hypothesis was about five times harder than the Poincaré conjecture, the only Millennium problem previously solved, by Grigori Perelman, after years of near-solitary work.
"It's just autocomplete" just got harder to say
For years the standard dismissal of large language models has been that they only predict the next word.
Technically, that's still true. Claude is trained to predict text.
But consider what that objection now has to explain. To predict the next step of a proof that has never been written down by anyone, there is nothing to copy. One widely shared joke captured it: the model is "just stochastically parroting what the person who solved the Riemann Hypothesis would say next." The punchline is that there was no such person.
The training objective is simple. What a system must learn to meet that objective at this level is not.
There's another difference people noticed. OpenAI's September result was described as a large-scale effort with many agents working in parallel, which invited the criticism that labs simply throw enormous compute at a problem until something falls out. Alpöge's note reads differently: weeks of one mathematician working closely with one model.
Either way, the conclusion is uncomfortable. Either very hard problems yield to methods simpler than we thought, or AI can now do genuinely creative mathematics. In both cases, work we assumed was safe from automation is not.
Mathematicians aren't finished. Their job is changing
It is tempting to read this as the end of mathematicians. Look closer at what happened.
Alpöge spent years preparing. He knew which tools mattered. He chose the framework. Claude accelerated past him at the end, but the setup was human.
Chess offers a partial analogy. Computers have beaten the best humans for decades, and professional chess still thrives. The difference is that in mathematics, the result matters to the rest of the world.
So the role shifts toward:
- Choosing the questions that are worth a machine's time.
- Judging which answers matter and how they connect.
- Turning verified results into understanding: shorter proofs, new concepts, teaching.
Alpöge says he expects the proof to be simplified dramatically, possibly quickly, and that he wants to understand and explain it properly. That's a mathematician's job description for the next decade in one sentence.
Knowing versus understanding
This is the genuinely new part.
A 250,000-page argument would take a dedicated human years just to read. Perelman's Poincaré proof filled a few papers. This one would fill a library shelf.
Tao recently wrote about almost exactly this scenario: a machine certifies a proof of the Riemann Hypothesis, and mathematicians are left not with doubt about the answer but with a different problem, the search for understanding. Certifying a result settles the question; it doesn't explain it.
For all of human history, knowing that something is true and knowing why have traveled together. They are starting to separate. And that won't stay inside mathematics. If AI can generate verified results faster than people can read them, the same gap will appear in chip design, materials and software.
The bigger implication: AI that improves AI
The reaction many researchers jumped to first wasn't about primes at all. It was about recursive self-improvement: AI helping to build better AI, which builds better AI again.
The argument goes like this. Much of AI research is a long list of well-defined problems: try an idea, measure it, keep what works. It rewards persistence and the ability to connect ideas from neighboring fields. Commentators such as Dwarkesh Patel have argued that this is exactly the kind of work current AI is disproportionately good at, and that it is even more valuable in machine learning than in pure math, because labs care about what makes training more efficient, not about theory for its own sake.
Now scale it. This result was reportedly one mathematician and one model. A lab can run thousands of copies of a strong model around the clock, and every improvement upgrades all of them at once.
Two caveats keep this grounded:
- Math is unusually checkable. Lean gives a clean right-or-wrong signal. Most AI research doesn't.
- Even optimists expect friction. Anthropic CEO Dario Amodei has said publicly that he expects transformative change to unfold over something like five to ten years, not hours, because human institutions absorb change slowly, and that barriers tend to fall "gradually and then all at once."
If anything, this result is an argument for the faster end of that range, because the hardest step in the loop, generating genuinely new ideas, just looked a lot less hard.
Is disease next?
In his 2024 essay Machines of Loving Grace, Amodei argued that powerful AI could compress 50 to 100 years of progress in biology and medicine into 5 to 10 years, including curing most infectious diseases and many cancers. Plenty of people called that science fiction. So was this, until Friday.
But biology is different in one decisive way: you can't compile a cancer drug.
In math, you write the proof and the checker tells you whether it's right. In medicine, you have to make the molecule, test it in cells, then animals, then people, win approval, manufacture it and deliver it. Every step takes time that no model can skip. Amodei himself has stressed that physical experiments, regulation and clinical trials set the pace once genius-level AI arrives.
So the realistic lesson is narrower and still big: domains with fast, reliable checks, like math, code and some chemistry simulations, will move first and fastest. Messy physical domains will speed up, but more slowly.
What about encryption?
Some people immediately asked whether this threatens internet security.
The Riemann Hypothesis itself does not break encryption. Modern public-key cryptography relies on problems like factoring huge numbers being hard, and proving where zeta zeros lie doesn't make factoring easy.
But the broader question is fair. Investor and writer Steve Newman asked publicly what the contingency plan is if a breakthrough on factoring comes next. That's a reasonable thing for security teams to think about, which is part of why the world is already migrating toward post-quantum cryptography standards.
What happens next
Watch for four things over the coming weeks:
- The remaining formal checks. The Comparator run, the paranoid kernel and a sandboxed rebuild.
- Expert review of the formal statement. Confirmation that the Lean theorem says exactly what mathematicians mean.
- A human-readable manuscript. Alpöge has said he will write one. Its quality will shape how quickly the field accepts the result.
- The Clay Mathematics Institute. Its prize rules require publication and a waiting period, so any formal recognition is far off.
FAQ
Has the Riemann Hypothesis officially been proved? Not officially. The Lean proof has passed many independent machine checks, but some checks were still running as of November 21, and the mathematical community hasn't formally accepted it.
Did Claude solve it alone? No. By Alpöge's account it was a six-week collaboration in which he set up the approach and a non-public version of Claude did much of the decisive work near the end.
Can I use the same Claude? According to the BBC, the model used is not publicly available.
Why release code instead of a paper? Because Lean lets computers check every step. The natural-language version would exceed 250,000 pages, which no human could practically review soon.
Does this break internet encryption? No. The hypothesis is about how primes are distributed, not about factoring large numbers.
The bottom line
A trained number theorist and an unreleased AI model say they have cracked mathematics' most famous problem in about six weeks, and they shipped the argument in a form that machines can check line by line. So far, independent checkers keep saying yes.
The caveats are real: unfinished checks, a narrower GRH than the headline suggests, a proof nobody can read in full, and a critic of Granville's stature calling the write-up garbled.
But even the cautious reading is staggering. Three years ago these systems fumbled arithmetic. Now the conversation is whether a machine has done what Riemann, Hilbert and generations after them could not, and the hardest part left for humans may be understanding the answer.
Sources
- Nature: Mathematician says he and AI solved Riemann hypothesis
- AP News: Anthropic worker claims to have solved a Holy Grail of math problems with the help of AI
- BBC News: Anthropic may have solved major maths problem
- Andrew Sutherland (MIT): Verification files for the Lean proof of the Riemann Hypothesis
- Science Monitor: Riemann Hypothesis Lean proof clears independent checkers
- Clay Mathematics Institute: The Riemann Hypothesis
- Dario Amodei: Machines of Loving Grace





