The Riemann hypothesis may have just fallen. Here is what it says.

On 20 November 2026, Levent Alpöge, a mathematician at Anthropic, announced a proof of the generalized Riemann hypothesis, and with it the Riemann hypothesis itself. The proof was produced with Anthropic’s Claude models. It comes as a four-page note and a Lean formalization of about 17 million lines.

The problem has been open since 1859. It is one of the seven Millennium Prize Problems and the one most mathematicians would pick as the most important open question in their field. So it is worth being precise about two things: what the hypothesis says, and where the claim stands today.

Where the claim stands

The short version is very promising, not yet confirmed.

Coverage: Nature, Scientific American, The Washington Post. I will update this note when the comparator verdict is in.

Now, the mathematics.

1735: a sum that knows about π

Add up the reciprocals of the squares:

$$1 + \frac{1}{4} + \frac{1}{9} + \frac{1}{16} + \cdots = \frac{\pi^2}{6}.$$

Euler proved this in 1735 and solved the Basel problem, which had beaten the Bernoullis. The general sum is the zeta function:

$$\zeta(s) = \sum_{n=1}^{\infty} \frac{1}{n^s}, \qquad s > 1.$$

1737: the same sum knows about primes

Two years later Euler found the identity that started analytic number theory:

$$\sum_{n=1}^{\infty} \frac{1}{n^s} = \prod_{p\ \mathrm{prime}} \frac{1}{1 - p^{-s}}.$$

The left side runs over all integers, the right side over the primes only. They agree because every integer factors into primes in exactly one way. That is the bridge: anything you learn about $\zeta$ is information about the primes.

You can watch the product creep up to $\pi^2/6$ with mathematicskit:

import math
from mathematicskit.special_functions import riemann_zeta
from mathematicskit.number_theory import euler_product

print(riemann_zeta(2.0), math.pi**2 / 6)   # 1.6449340668... twice
for N in (10, 100, 10_000):
    print(N, euler_product(2.0, N))        # 1.5951, 1.6419, 1.644918

Ten primes’ worth of product already gets two digits right. The Basel and Euler product gallery example plots the convergence.

1859: Riemann goes complex

In an eight-page paper, his only one on number theory, Bernhard Riemann did three things:

  1. He let $s$ be a complex number and extended $\zeta(s)$ to the whole plane except $s = 1$, by analytic continuation.
  2. He found a functional equation relating $\zeta(s)$ to $\zeta(1-s)$, a mirror symmetry about the line $\mathrm{Re}\ s = 1/2$.
  3. He wrote down an explicit formula for counting primes as a sum over the zeros of $\zeta$.

The continuation gives famous values that the original series cannot: $\zeta(0) = -1/2$, $\zeta(-1) = -1/12$, and $\zeta(-2) = \zeta(-4) = \zeta(-6) = \cdots = 0$. Those last ones are the trivial zeros:

import numpy as np

print(riemann_zeta(0.0), riemann_zeta(-1.0))          # -0.5  -0.08333...
print(riemann_zeta(np.array([-2.0, -4.0, -6.0])))     # [0. 0. 0.]

All the other zeros, the nontrivial ones, lie in the strip $0 < \mathrm{Re}\ s < 1$. The first is at $s \approx \tfrac12 + 14.1347\ i$.

The hypothesis

Every nontrivial zero of $\zeta(s)$ has real part exactly $\tfrac12$.

Riemann wrote that this was “very probable” and that he had put the search for a proof aside. Since then, more than ten trillion zeros have been computed, and every one sits on the line $\mathrm{Re}\ s = \tfrac12$. Numerical evidence on that scale does not prove anything about the infinitely many zeros beyond, which is why the problem has been open so long.

Why primes care

Let $\pi(x)$ count the primes up to $x$. Gauss guessed that $\pi(x)$ is close to the logarithmic integral $\operatorname{li}(x) = \int_0^x dt/\ln t$. The prime number theorem (1896) proved that the ratio tends to 1. It did not say how big the error is.

The zeros of $\zeta$ control that error. In Riemann’s explicit formula, each zero $\rho$ contributes a wave of size about $x^{\mathrm{Re}\ \rho}$. Zeros on the critical line give waves of size $\sqrt{x}$, and von Koch showed in 1901 that RH is equivalent to

$$|\pi(x) - \operatorname{li}(x)| \le C\sqrt{x}\ \ln x.$$

In words, the primes are as regular as they can possibly be. They are as random as coin flips, and no more.

from mathematicskit.number_theory import prime_counting, logarithmic_integral

x = np.array([10**3, 10**4, 10**5, 10**6])
print(prime_counting(x))          # [  168  1229  9592 78498]
print(logarithmic_integral(x))    # [  177.6  1246.1  9629.8  78627.5]
print(x / np.log(x))              # [  144.8  1085.7  8685.9  72382.4]

At a million, $\operatorname{li}(x)$ is off by 130 primes, while the naive $x/\ln x$ is off by more than 6,000. The prime number theorem gallery example draws the race.

Hundreds of published theorems begin “assuming RH…” or “assuming GRH…”. They cover how quickly a primality test runs, the size of the least quadratic non-residue, and class numbers. If the proof holds, all of them become unconditional at once. That is why this claim matters even to people who never think about $\zeta$.

What to watch

The mathematics will not be settled by press releases. It will be settled by three things:

Each of these can be checked in public. That is new, and whatever the verdict, it is how big claims in mathematics will be tested from now on.

For the history up to the eve of this announcement, see the special functions history page and the number theory history page of mathematicskit, and the Riemann zeta gallery example.