A four-page note made the claim; 239,275,554 bytes of Lean 4 code carried the argument. On Friday, Anthropic mathematician Levent Alpöge posted the note on his personal website, saying he had proved the Riemann hypothesis with Claude, Anthropic’s AI model, The New York Times reported. A technical inventory by Lean 4 Dev describes a PDF dated 20 November 2026 and a 14-part archive totaling 239,275,554 bytes, or about 239 MB using decimal units. That is the size of the claimed formal development, not evidence that its mathematics is correct.

Alpöge’s own wording was deliberately short of a declaration that verification was complete: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” Al Jazeera reported that mathematicians were treating the Claude-assisted claim as potentially historic. The important distinction is between a proposed proof and one whose statement and derivation have both survived independent checking.

The Riemann hypothesis concerns the zeros of the zeta function. For complex values of s with real part greater than 1, that function can be written as ζ(s) = 1 + 1/2^s + 1/3^s + …, and its definition extends beyond that initial domain. Its nontrivial zeros are the zeros other than the negative even integers. The hypothesis says every nontrivial zero has real part exactly 1/2: they all lie on one vertical line in the complex plane.

Translucent coral and pale teal waves cross a textured teal background behind a vertical coral line dotted with dark circles.

Alpöge also says his work establishes the Generalized Riemann Hypothesis, which makes the corresponding claim for a wider class of functions, Scientific American reported. That makes the final theorem statement especially important. A shorter informal description can hide exactly which objects, definitions and quantifiers a formal proof covers.

What compiling a Lean proof establishes

Lean is a formal proof system. Instead of reading a mathematical argument and judging whether each step follows, a checker processes formal definitions, a theorem statement and proof terms. A successful run means the submitted derivation type-checks under Lean’s rules and whatever definitions and assumptions the development includes. It removes the ordinary possibility that a reader overlooked an invalid algebraic or logical step. It does not, by itself, establish that the English name attached to the theorem matches the mathematics encoded in Lean.

MIT mathematician Andrew Sutherland and others independently compiled Alpöge’s proof code, MLQ.ai reported. Sutherland put the probability that “anything is fishy” at “approximately zero.” That is an expert judgment based on the checks performed, not a measured failure rate. At the time of that report, the separate check that the formal statement matched the Riemann hypothesis was still pending.

A vintage-style illustration shows a machine processing stacks of paper, with a magnifying glass highlighting the sheets.

Those are two different tests with two different failure modes. Suppose the code compiles but the encoded statement is weaker than intended: Lean has then certified a valid proof of the wrong target. Conversely, an accurately encoded target attached to a broken derivation will not produce a valid machine-checked proof. The independent compilation addresses the second failure mode much more directly than the first.

What is still unmeasured

The initial reporting also supplies no quantitative division of labour between Alpöge and Claude. Calling the result Claude-assisted records Alpöge’s account of the collaboration; it does not show whether the model supplied the main construction, individual lemmas, routine formalisation or some combination of them. A compiling Lean archive verifies a mathematical artifact, not an attribution model for who produced each part of it.

Scientific American noted that issues can arise in Lean but that experts should be able to check whether the hypothesis was encoded properly and whether the proof is correct. The next result to watch is therefore binary rather than statistical: an independent statement check either matches the final Lean theorem to the Riemann hypothesis, including its definitions and quantifiers, or it does not.