Key points:
• Anthropic mathematician Levent Alpöge has claimed a Lean-formalized proof of the Riemann Hypothesis and Generalized Riemann Hypothesis, done in collaboration with Claude in about 41 days.
• Independent verifications by MIT’s Andrew Sutherland and other researchers have found that the 175,000+ module archive passes compilation without any standard axioms and is accepted by many external verifiers.
• This breakthrough claim, if accepted completely, will solve a 167-year old Millennium Prize problem and mark a move towards AI-powered formal discovery where the human insight is yet to come out.
On the morning of November 20, one day after the Anthropic IPO, Levent Alpöge published a four-page note on their website stating that they and Claude have proved the Riemann Hypothesis (proposed by Bernhard Riemann in 1859) states that all nontrivial zeros of the zeta function lie on the line Re(z)=1/2. As Alpöge mentions, these results generalise to the generalised Riemann Hypothesis for Dirichlet L-functions.
No paper was published in the usual sense. Instead, a Lean proof of the result was published, spanning around 175187 files and over 17 MILLION lines of code. This is equivalent to a hand written manuscript of over 250000 pages, which Alpöge says they chose not to publish due to being “slop.”
Alpöge, a PhD student at Princeton under Manjul Bhargava, and junior fellow at Harvard, has collaborated with AI to produce multiple results this year. In this case, Alpöge set the overall direction, with Claude providing key steps towards the end of the proof. “Claude was much faster. […] Near the end it was really fast, I just watched them fly and dance and sparkle.”
Read:
- How Learning New Skills Builds Confidence and Personal Growth
- Technology Trends Every Educator Should Know in 2026
The note covers work from early October up to the 16th of November, utilising methods such as Stepanov’s method, theta lifting, along with other techniques from arithmetic geometry and automorphic form theory. A message was sent to Anthropic CEO Dario Amodei before the result was made public. Alpöge comments that making the result public so soon after the IPO was “really bad timing”.
Verification has proceeded exceptionally quickly. The repository has been independently built by Andrew Sutherland (MIT). Full verification of the exported proof has been performed by multiple people (nanoda, con-leche, lean4lean ). All five main theorems were proven using only the standard 3 axiomatic assumptions within Lean. Sutherland comments that there is effectively a zero chance of this result being “fishy”.
Comparisons and paranoid mode verification are still being carried out by other Lean experts. While we have very strong evidence that the theorem follows from the axioms through verification by a trusted kernel, we cannot be sure that the mathematical community understands why the proof works. Alpöge indicates that we will see a proper explanation by humans at some point, but as it stands, mathematicians must work backwards from a machine proof of the result.
This is not the first result from AI on this problem. In August, an unpublished result from Claude increased the known proportion of zeros on the critical line to 67.2%, from 41.6%. If this result is true, it means hundreds of conditional theorems can now be considered unconditional. Furthermore, it gives great evidence that, through formal verification, large language models can now complete arguments so significant that careers have been built around them.
However, the difficulty of digesting a quarter of a million page collaboration makes it difficult to match the traditional idea that proofs should be understandable to an expert human. Increasingly, strategy and interpretation of key ideas may be done by mathematicians, while formal proofs of significant size are produced by, and verified by machines.

