20 November 2026

How deep is the Riemann hypothesis?

Today Levent Alpöge, a mathematician at Anthropic, said that he and the company's AI had proved the Riemann hypothesis, the most famous open problem in mathematics. Nobody has confirmed it yet, and it may not survive checking. Most of the coverage skips over what the problem actually says, so this page is my attempt to explain it. It takes about ten minutes, and most of it you can play with.

Scroll down. The further you go, the more school it takes.
Start here

Primes look random, but only up close

A prime is a number bigger than 1 that can't be split into equal piles, apart from piles of one. Six can be two piles of three, but seven only ever comes out as seven. There is an old method for finding primes, called a sieve. Write out the numbers from 2 up, keep 2 but cross out every other multiple of 2, then do the same for 3, 5 and 7. Whatever survives below 100 is prime.

25 primes below 100.

Now look at where they fall. The gaps between neighbouring primes run 1, 2, 2, 4, 2, 4, 2, 4, 6, 2 and carry on in that vein, without any rhythm you could tap out. People have been looking for a rule behind the gaps for more than two thousand years, and nobody has found one.

Step back far enough, though, and the primes turn out to be surprisingly well behaved. Around a million, roughly 1 number in 14 is prime. Around a trillion it is 1 in 28. They thin out at a steady, predictable rate.

So from a distance the primes follow a schedule, and up close they are a mess. The Riemann hypothesis is a statement about that mess. It says exactly how large the mess is allowed to get.

Bernhard Riemann, 1859

A smooth estimate, and the waves that correct it

Imagine walking along the number line and keeping a tally of the primes you pass. The tally goes up by one at every prime, so if you draw it you get a staircase. That is the dark line in the chart below.

The red line is an estimate of the same tally, worked out from the rate at which primes thin out. It is a good estimate, but it is always a little off, because a smooth curve has no way of making steps.

In 1859 Bernhard Riemann published a way to correct it, in the only paper he ever wrote about primes. His correction is a list of waves. Each wave pushes the estimate up in some places and down in others. Add them one at a time with the slider and watch the red line start to find the steps.

The real tally of primesThe estimate, plus waves
Waves added: 12Average miss: 0 primes
The newest wave on its own.

A hundred waves get you most of the staircase. The list of waves goes on for ever, and if you could add them all, the red line would match the staircase exactly. That much has been proved, and it is true whether or not the hypothesis is.

Each wave is fixed by a single number. Riemann found those numbers inside a formula called the zeta function. They are the places where the formula comes out as zero, which is why they are called its zeros. Taken together, the zeros are a recipe for the primes.

The actual mathematics

The formula behind the waves

This section has real formulas in it, but only three steps' worth, and each step comes with something to play with. By the end you will be able to state the hypothesis the way a mathematician would.

1. An endless sum

Pick a number and call it s. Start with 1, add 1 over 2 to the power s, then 1 over 3 to the power s, and keep going without end. As long as s is bigger than 1, the running total settles down to a fixed value. That value is written ζ(s), read “zeta of s”.

ζ(s) = 1 + 12s + 13s + 14s + 15s + …

2. Euler's surprise: the same value, built from primes

In 1737 Leonhard Euler found a completely different route to the same value. You multiply instead of adding, and you use only the primes.

ζ(s) = 11 − 2−s × 11 − 3−s × 11 − 5−s × …

It works because every whole number can be built from primes in exactly one way. If you multiply all those brackets out, each whole number turns up exactly once.

Adding, over all numbersMultiplying, over primes only
s = 2.0ζ(s) = 1.6449

This is why primes come into it at all. One side of the equation involves every whole number, and the other involves only the primes. That means ζ carries information about where the primes are, and whatever you manage to learn about ζ tells you something about them.

3. Riemann's move: numbers with two coordinates

Riemann's idea was to let s be a complex number. A complex number has two parts, called its real part and its imaginary part, so you can picture it as a point on a map. The real part says how far across the point is, and the imaginary part says how high. The endless sum breaks down for points to the left of 1, but there turns out to be exactly one smooth way to extend the function over the whole map, and Riemann found it.

On this larger map, ζ sometimes comes out as exactly zero. You can watch it happen. Move straight up the map with the slider and follow where ζ lands on the right. The cross marks zero.

Walk up at real part:
Height: 28.00

At real part ½ the path keeps swinging back through the cross, and every time it does, that is a zero. Now try ¼ or ¾. The path loops close to the cross but never touches it, and that has been true as far up as anyone has looked.

The Riemann hypothesis, stated properly
Every zero of ζ(s) whose real part lies between 0 and 1 has real part exactly ½.

ζ has other zeros, at −2, −4, −6 and so on. Those are easy to find and don't count.

How a zero becomes a wave

Start from Euler's product and ask how many primes there are below some number x. Riemann showed that the answer is the smooth estimate, minus one term for every zero. For a zero with real part ½ and imaginary part t, the term is built from x½ + it, and that splits into two parts:

x½ + it = √x × a wave in ln x, running at speed t

The imaginary part of the zero controls how fast the wave wiggles. The real part controls how fast the wave grows: ½ gives growth like √x, and anything larger than ½ would grow faster. These are the same waves you were adding to the chart earlier.

The hypothesis itself

Riemann's bet

So each wave is shaped by the two parts of its zero. The imaginary part sets how tightly the wave wiggles, and it is different for every wave: 14.13 for the first, 21.02 for the second, and upward without end. The real part sets how quickly the wave grows as the numbers get bigger.

Riemann wrote that it was “very probable” that the real part is exactly ½ for every zero. He added that he had made a few brief attempts to prove it, hadn't managed, and was putting it aside because he didn't need it for what came next. Then he went on with the paper.

That passing remark is the Riemann hypothesis. If you plot the zeros on the map, the bet is that all of them sit on a single vertical line. The picture shows the first 29.

It matters because a zero with real part above ½ would give a wave that grows faster than all the others and eventually drowns them out. The estimate would then miss by more, and the primes would be less orderly than everyone believes.

To put numbers on it, take the primes below a trillion. There are 37,607,912,018 of them, and the simplest smooth estimate is off by only 38,263. If Riemann was right, the miss at that point could never have been more than about a million. If he was wrong, there is no such guarantee.

Computers have checked the first twelve trillion waves, and every one is on the line. Unfortunately that proves nothing, because the list is infinite, and number theory has examples of patterns that hold for far longer than that before they fail.

There is also a stranger reason people believe it. In 1972 Hugh Montgomery worked out how the imaginary parts of the zeros are spaced. Over tea in Princeton he showed the pattern to the physicist Freeman Dyson, who recognised it straight away: the energy levels of heavy atomic nuclei are spaced in the same way. Nobody has explained why.

Progress on paper has been slow. Each bar below is the share of waves that somebody had proved to be on the line by that year.

1914G. H. Hardy: infinitely many, which could still be a share of zero
1942Atle Selberg: a real share, though a tiny one
1974Norman Levinson: more than a third
1989Brian Conrey: more than two fifths
2020Four authors, thirty-one years later: five twelfths
2026August, with AI help: about two thirds. Not yet peer reviewed.
2026Today's claim: all of them.

Striped bars are announcements that have not been through peer review.

Hardy, whose result starts that chart, once sent a postcard to a friend before a rough sea crossing, saying that he had proved the Riemann hypothesis. He hadn't. His reasoning was that God would never let him drown and be remembered for that.

Going down

What it takes to work on it

What you have seen so far is the cartoon version. Working on the real problem takes about twenty-two years of education, and people drop out at every stage along the way.

Each band below is one stage. The bar shows roughly how many people alive today have got that far. The scale is compressed, so every tenth of the bar's length stands for a factor of ten in people.

The head counts are my own rough guesses, so please don't quote them.

Those few hundred people are everyone who was in a position to try.

Meanwhile

The theorems that were waiting

Mathematicians did not wait for a proof. For about a century they have been publishing results that begin “assuming the Riemann hypothesis”. Many of those results need a broader version of the bet, and today's claim is said to cover that version too, so it is worth explaining first.

The generalized Riemann hypothesis

The zeta function treats all primes alike, but you can ask finer questions. Leaving out 2, every prime is either one more than a multiple of 4 (like 5, 13 and 17) or three more (like 3, 7 and 11). Call these team 1 and team 3. You might wonder whether the primes prefer one team.

To study that, you adjust the endless sum. Drop the even numbers and give the team-3 numbers a minus sign:

L(s) = 1 − 13s + 15s − 17s + 19s − …

This is a relative of zeta called an L-function. It has its own Euler product, its own zeros and its own waves, and those waves govern the race between the two teams. There is one of these relatives for every way of sorting numbers by their remainders, whether by last digit, by remainder after dividing by 7, or anything else. So there are infinitely many of them.

The generalized Riemann hypothesis makes the same bet for all of them at once: every zero with real part between 0 and 1 has real part exactly ½. In plain terms, primes are shared out evenly among the teams, and no team ever gets ahead by much more than a square root.

Team 3's lead over team 1Level
Count primes up to:

Team 3 is almost always slightly ahead. Team 1 first takes the lead at 26,861 and loses it again immediately. Even so, the lead stays tiny compared with the thousands of primes on each side, and the generalized hypothesis says it always will.

You will also come across the name “extended Riemann hypothesis”. Some writers use it for the same bet about number systems larger than the whole numbers, such as the ones you get by adding √2 or the square root of −1. Others use the two names interchangeably. Press reports say today's claim covers the generalized version. I have not checked which family of functions the code's statement actually includes.

What was waiting

Here are a few of those results. Flip the switch to see which of them would lose their “if”.

Only if the bet is rightWould be proved

A test for primes with no luck involved

In 1976 Gary Miller found a fast way to tell whether a number is prime, but it is only guaranteed to work if the generalized hypothesis is true. A slower method with no conditions attached was found in 2002.

Only if the bet is rightWould be proved

No long deserts between primes

After a prime p, you never have to wait longer than about √p × ln p for the next one. Harald Cramér showed this in 1920, assuming the hypothesis. The true wait is probably much shorter.

Only if the bet is rightWould be proved

How often a fraction behaves like 1/7

1/7 is 0.142857 142857…, repeating every six digits, which is the longest a seventh can manage. If the generalized hypothesis holds, about 37% of primes behave the same way. Christopher Hooley, 1967.

Only if the bet is rightWould be proved

A question Ramanujan left open

Which odd numbers can't be written as x² + y² + 10z²? In 1997 Ken Ono and Kannan Soundararajan gave the complete list, assuming the generalized hypothesis. The largest is 2,719.

Only if the bet is rightWould be proved

Euler's 65 numbers

Euler found 65 “convenient numbers” that helped him search for large primes. In 250 years nobody has found a 66th, and if the generalized hypothesis holds, there isn't one.

One that didn't wait

Every odd number above 5 is three primes added together

Proved in 1997 with the “if” attached. In 2013 Harald Helfgott removed the “if” by proving it the hard way.

A common mix-up

It doesn't break encryption

The hypothesis is about where the primes are. Breaking encryption would need a fast way to split a huge number into its prime factors, which is a different problem.

The bottom

175,187 files

The proof did not arrive as a paper. It arrived as computer code in a language called Lean, in which a program checks every logical step and rejects anything that doesn't follow. Press reports put it at about 175,187 files. There is a four-page note explaining the idea, but nobody has yet written out the full argument for people to read.

Below, each dot is one file. Files vary a lot in length, so this shows how many there are and not how much is in them.

If you read one file every five minutes, eight hours a day with no weekends off, you would finish in 2031.

So is it proved? · As of 20 November 2026

  • Not yet. Officially the problem is still open.
  • Who says it's proved? Levent Alpöge, a mathematician at the AI company Anthropic. He says he found the proof while working with its AI system, Claude.
  • Has anyone checked? Only the first steps. Press reports say an independent group confirmed that the files match what was published and that the main theorem statements compile. Nobody outside has yet confirmed that the whole proof checks, or that the code states the hypothesis the usual way. I have not seen any of this first hand.
  • What's missing? Peer review, a full version a person can read, and a careful audit of the definitions the code starts from.
  • Anything else? A second proof, by Alessandro Masullo and written without AI, was reported the same day. It is unconfirmed as well.
  • The prize? One million dollars, but only after publication and two years without anyone finding a hole.

For 167 years, the people at the bottom of that descent couldn't do it.

If this holds up, the first proof will be one that none of them has been able to read yet.