Claude Riemann Hypothesis Result: 41.6% to 67.2%

Anthropic pointed an unreleased research version of Claude at the Riemann hypothesis, and the model came back with a better number than the one theorists had been inching forward for half a century: the proven fraction of the zeta function's zeros that satisfy the hypothesis rose from 41.6% to 67.2%. The company posted that result on August 10, 2026, said plainly that Claude did not solve the hypothesis, and pointed to a research write-up on its own site. The announcement is sitting at 8.7 million views.

"We asked an unreleased research version of Claude to take a stab at the Riemann hypothesis. It didn't solve it, but it did make strides on a related problem: it increased the lower bound for the fraction of zeros of the Riemann zeta function that satisfy the hypothesis from 41.6% to 67.2%." Anthropic, on X

There is no setting in your account that turns any of this on, and the claim is mathematical before it is commercial. So start with the mathematics.

Count what the 67.2% is a fraction of

The Riemann zeta function is the thing you get when you take the sum 1 + 1/2^s + 1/3^s + 1/4^s and extend it to complex values of s. It has zeros. Some are boring and sit at the negative even integers. The rest, called the nontrivial zeros, are trapped inside a vertical strip, and Riemann's conjecture says every single one of them has real part exactly 1/2. Draw that as a vertical line up the complex plane and you have the critical line. The hypothesis is the assertion that no zero ever wanders off it, and nobody has proved it.

There are infinitely many nontrivial zeros, so "fraction" needs a working definition. The standard one: count all the nontrivial zeros up to some height T, count how many of those lie on the critical line, take the ratio, then watch what that ratio does as T grows without limit. If the hypothesis is true, the ratio is 1 forever. Every result anyone has proven is a floor underneath that ratio. Both 41.6% and 67.2% are floors, not measurements.

A floor cuts one way only. The 32.8% left over are not zeros known to sit off the critical line. They are zeros that no argument has been able to place at all, in either direction. Raising the floor from 41.6% to 67.2% shrinks that unplaced remainder from a majority of the zeros to a minority of them, and it tells you nothing new about where any individual zero is. A single zero found off the line would end the hypothesis on the spot, and no such zero appears anywhere in Anthropic's announcement. The number moved because the proof got stronger.

Watch how a floor like this gets pushed up

Results of this shape come from a standard piece of machinery in analytic number theory, and the general outline explains why the figures usually move by fractions of a point. You multiply zeta by a mollifier, a short Dirichlet polynomial engineered to flatten out zeta's wild fluctuations in size. Once the product behaves, a criterion converts an averaged estimate of the mollified function into a count of zeros sitting on the line. The proportion that falls out is governed by how long you are allowed to make that mollifier. Longer mollifier, better proportion, much nastier mean-value estimate underneath it. Every gain is paid for in analytic difficulty, which is why work on this bound tends to show up as small increments rather than jumps.

Then the arithmetic that made mathematicians look twice. The bound Claude improved is roughly fifty years old, and 41.6% is where it stood when Anthropic started. This one result claims 25.6 percentage points in a single step. A jump that size is either a genuine change in method or an error somewhere.

Separate a density result from a proof

Both numbers in this story come from Anthropic, measuring its own model's output, on a problem Anthropic chose. Anthropic's post itself carries the two figures, one sentence of framing and a link to a research write-up on anthropic.com. The write-ups that followed within a day, on DataCamp and Neowin among others, describe the run as a multi-agent Claude session whose resulting argument was formalized in Lean, the proof assistant that machine-checks each individual step of a mathematical argument.

A Lean file compiles or it does not. If a formalization covers the whole argument, and the statement at the top of the file says what the English summary says it says, then the result stands on its own and no part of believing it depends on trusting Anthropic or the model that wrote it. Those two conditions are exactly where a careful reader should push. Whether the formalization has been released in a form outside mathematicians can compile is not established by the announcement. A model that produces a plausible-looking proof is a demo. A model that produces a compilable one has taken itself out of the trust chain entirely, which is a far stronger position than any benchmark score.

Now the ceiling, because the headline hides it. Pushing this bound to 100% still would not prove the Riemann hypothesis. Proving that the proportion tends to 1 leaves room for infinitely many rogue zeros off the line, provided they thin out fast enough to be a vanishing share of the total. A proportion is a statement about density. The hypothesis is a statement about every zero without exception, and the two targets never quite meet. Anthropic's own wording was accurate. The post says strides on a related problem, and the related problem is not the hypothesis.

Log into the Claude that actually exists

The model that did this is a research version with no public identity. What you can sign into is the Claude that ships today, and no prompt you write in it reproduces a result that took a dedicated multi-agent research run to find. Treating the announcement as a description of your own tool's abilities will only mislead you.

What the result does tell you about using Claude on mathematics is more useful than the percentage. The shape that worked here is a search process bolted to a checker: generate candidate arguments fast, formally verify each one, discard everything that fails to compile. You can run a small version of that pattern today. Give Claude a statement you can independently verify rather than one you have to accept, use it against symbolic computation and literature retrieval where a wrong answer surfaces immediately, and keep a verification step outside the model in the loop. The verification step is what turned an interesting output into a claim mathematicians are willing to read.

Read the research write-up linked from the announcement before forming a view, and weigh the Lean statement above the English summary wherever you can see both. The search that produced this ran once, inside Anthropic, on a model with no version number, and how many attempts it took to land on 67.2% is a count only Anthropic holds.