AI may have just proved the Riemann hypothesis
An Anthropic mathematician says he and an unreleased Claude model have proved the most famous unsolved problem in maths. Computers have checked almost all of the proof, but no other mathematician has worked through it yet.
This is my seventh day in a row posting, and AI isn’t slowing down. On Friday, Levent Alpöge, a mathematician at Anthropic, said he and an unreleased Claude model had proved the Riemann hypothesis. It’s been open since 1859 and has a $1m prize on it.
The market dropped anyway. The Nasdaq fell 5.3% that day, and Anthropic’s own shares, on only their second day of trading, fell 9.1%.
The hypothesis is about where the zeros of one function, the Riemann zeta function, sit. Riemann guessed that the important ones all sit on a single line. If he was right, prime numbers are spread out about as evenly as they possibly could be, and a lot of number theory has been built on assuming he was.
The proof is 17.4 million lines of Lean, a programming language that lets a computer check every step. Alpöge says a version written out in English ran past 250,000 pages, so he didn’t publish one: “It being obviously inappropriate to attempt to post some massive unreadable travesty of ‘slop’, I decided to only present Lean.”
Where the checking standsFrom MIT’s verification page, last updated 21 November, 09:30 GMT
- All 175,187 files of the proof compile Passed
- It relies only on Lean’s three standard axioms Passed
- Three proof checkers written by other people accept it, as does Lean’s own replay Passed
- It proves the hypothesis as Mathlib, Lean’s shared maths library, already defined it Passed
- The final end-to-end comparison and Lean’s strictest kernel checks Running
- A rebuild in a sealed sandbox Not yet
- A version a mathematician can read Not yet
- A journal paper, then two years before the $1m prize Not yet
Andrew Sutherland at MIT, who is running the checks, told Scientific American: “Honestly the chance that anything is fishy here is approximately zero.” No other mathematician has worked through the argument yet, though. Terence Tao expects “the digestion process will take some time”, and the Clay Institute, which awards the prize, says assessing it “takes time”. Its rules need the proof published in a journal and then two years to pass.
We’ll update this post when MIT’s last checks finish.
Faster than the mathematicians expected
In May last year Daniel Litt, a mathematician at Toronto who follows AI closely, wrote that it was “exceedingly unlikely AI will prove the Riemann Hypothesis in 2026”. On Friday he posted: “Seems like I was wrong about this!”
In October OpenAI proved a weaker version called the quasi-Riemann hypothesis, which rules out zeros in part of the area where they could be. Alex Kontorovich, a number theorist at Rutgers, went on Lex Fridman’s podcast to talk about it.
“I did not think I would see a proof of the quasi-Riemann hypothesis in my lifetime.”
Alex Kontorovich on the Lex Fridman Podcast, episode 503
He also said, “I’m already impressed with quasi-RH, but it’s not RH.” The episode came out on 26 October, 25 days before Alpöge’s note. On Friday he posted that he was “a bit more confident that the proof has a good chance”, because it uses “everything”.
Eighteen months of AI mathsResults in white, the prediction in grey
- May 2025 Daniel Litt: “exceedingly unlikely AI will prove the Riemann Hypothesis in 2026”
- Jul 2025 OpenAI and Google DeepMind models score gold at the International Mathematical Olympiad, the world’s top maths contest for school students
- May 2026 An OpenAI model disproves Erdős’s unit distance conjecture. Thomas Bloom: “the most impressive achievement of AI in mathematics so far”
- Jul 2026 Claude and two mathematicians find a counterexample to the Jacobian conjecture, open since 1939
- Sep 2026 Claude writes Fermat’s Last Theorem out in Lean: 13 million lines
- Sep 2026 OpenAI claims Navier–Stokes, one of the seven Millennium Prize problems (disputed)
- Sep 2026 AI scores 98% on FrontierMath’s hardest tier, up from 5% fourteen months earlier
- Oct 2026 OpenAI proves the quasi-Riemann hypothesis
- Nov 2026 Levent Alpöge and Claude claim the Riemann hypothesis
Most people haven’t noticed
I was talking to a friend the other day about the Navier–Stokes result, and he hadn’t heard of it. In September OpenAI said its AI had solved one of the seven $1m Millennium Prize problems, about whether the equations for flowing water and air can break down. Many mathematicians said it relied on a loophole in how the problem is worded. OpenAI’s post about it got 75 million views on X.
I looked at the Wikipedia page views for each problem. Each time a result came out, views of the problem’s page went past 100,000 a day. Within a week they’d dropped by more than 90%.
Wikipedia views of each problem’s page, from the day before the newsDaily views, all three on the same scale
Jacobian conjecture20 Jul
Navier–Stokes8 Sep
Riemann hypothesis20 Nov
See the numbers
| Page | Day before | News day | Peak | 7 days later |
|---|---|---|---|---|
| Jacobian conjecture | 28 | 184,222 | 184,222 | 6,315 |
| Navier–Stokes existence and smoothness | 1,109 | 76,527 | 105,389 | 6,005 |
| Navier–Stokes equations | 2,168 | 110,845 | 153,538 | 10,841 |
| Riemann hypothesis | 964 | 295,157 | 295,157 so far | Not yet |
The Riemann hypothesis page had 295,000 views on Friday, against about 1,000 on a normal day, and the news only broke in the afternoon UK time. That’s more than either of the others managed on any day.
Andrej Karpathy, who co-founded OpenAI, thinks the gap in understanding of what AI can do is “widening”. His napkin maths:
Share of the world’s populationAndrej Karpathy’s estimates
The market didn’t care either
Anthropic sold its shares at $83 when it listed on the Nasdaq on Thursday, and closed its first day at $97.17. Alpöge’s note went up on Friday afternoon UK time, around when Wall Street opened. By the close, Anthropic was at $88.30 and Nvidia had lost 13.8%.
The market reports put the selling down to worries about AI valuations and chip stocks. None of the ones I read mentioned the Riemann hypothesis.
Andrew Granville, a number theorist in Montréal, told Nature the release was “garbled” and linked its timing to the share price. Alpöge saw that coming in his note: “There is an obvious cynical view, to do with the timing, but it is false.”
What it means for encryption
Not much, directly. Most encryption on the internet depends on one thing being hard: taking a very large number and finding the two primes that were multiplied together to make it. On the same podcast, Kontorovich imagined one of OpenAI’s 700 October papers turning out to be a fast way to do that.
“Everybody would know. Like, you can’t use your credit card on Amazon anymore. All of e-commerce is gone.”
Alex Kontorovich on the Lex Fridman Podcast, episode 503
A proof of the Riemann hypothesis doesn’t do that. Cryptographers have mostly assumed it was true for years.
What AI has found in encryption this year
- Legacy 1024-bit RSA keysGoogle DeepMind, 6 Nov
- AI agents working with cryptographers found ways to factor them in bulk. Google says modern 2048-bit keys aren’t affected.
- AES, cut down to 7 of its 10 roundsAnthropic, 28 Jul
- Claude Mythos Preview found an attack 200 to 800 times faster than the previous best. Full AES isn’t broken, and Anthropic says no systems in use are affected.
- HAWK, a candidate post-quantum signatureAnthropic, 28 Jul
- Claude found an attack that halves its effective key size. Breaking the smallest version dropped from about 2⁶⁴ steps to 2³⁸. It isn’t in use yet.
- The Riemann hypothesis20 Nov
- No direct effect. Encryption doesn’t depend on it, and most cryptographers already assumed it was true.
Google DeepMind’s conclusion was that “we can no longer rely on historical hardness assumptions and must continuously stress-test cryptographic standards.” The UK’s National Cyber Security Centre already wants organisations to know what encryption they use by 2028, to have moved their most important systems to post-quantum methods by 2031 and to have finished by 2035.
If anything you run still uses 1024-bit RSA keys, Google has now shown a way to break them in bulk.