The message fits on one line. Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, describe a problem to Claude: compute a certain particle physics formula "at nine loops".
Then they add, in essence: "I'm going to sleep and won't be available for several hours. Keep working on this until I tell you to stop. Give me updates every four to six hours."
And they go to bed.
A week later, the calculation is done. Until then, the record stood at eight loops.
The most fragile calculation in the world
To predict what happens when particles collide, for example in CERN's great accelerator, physicists use formulas called scattering amplitudes. They tell you how likely the particles are to react in one way or another.
The problem is that these formulas cannot be computed in one go. You proceed by successive layers of corrections, called "loops". Each loop brings you closer to the true result, a bit like each extra decimal brings you closer to the true value of a measurement. But each loop is also far harder than the one before.
Most calculations stop at two loops. The most precise prediction in all of particle physics, the one in the textbooks, uses five.
In 2023, Lance Dixon, a professor at Stanford University's SLAC laboratory, reached eight loops for a benchmark case used by theorists. He thought the ninth would be too hard to compute directly.
One week, 96 processors, a sudoku
Claude used a method called the "bootstrap". The idea is like a sudoku: rather than working out the answer square by square, you write down every possible answer, then eliminate those that break the known rules. In the end, only one is left.
That part used 96 processors for a week, for about a hundred dollars. In total, counting the cost of running Claude itself all that time, the operation would have cost a customer "around one or two thousand dollars", according to the account Anthropic published on 25 September.
Above all, Claude did the calculation two different ways: with the bootstrap, then with the indirect method Lance Dixon had used for his record. Both paths lead to the same result.
The judge is the record holder
A result like this is worth nothing until someone has checked it. And the best person to do so was Lance Dixon himself.
He checked the work. His verdict: Claude used "all the methods my collaborators and I developed over the years".
He also stresses what makes the feat remarkable. In this kind of calculation, he says, "if you make any mistake at all in the computational recipe, it all crashes down like a failed soufflé". A week of unsupervised work, thousands of steps, and not a single fatal error.
During the checking, the physicists learned that a team at the Chinese Academy of Sciences, led by Song He, had already obtained most of the same result. It had used an AI based on GPT-6, but in a very different way: the researchers kept control of their own method and handed part of the calculations to the machine. Two teams, two ways of working with AI, the same record broken almost at the same time.
What it changes, and what it doesn't
It is worth being precise about what happened. Claude did not invent a new theory. It applied, with remarkable rigour, methods that humans took years to build.
But that is already a lot. For a week, with nobody watching, it carried out work where the slightest mistake ruins everything. It is exactly the kind of task that used to go to a brilliant PhD student, for months.
For science, the question shifts. If a machine can produce this kind of result in a week, the bottleneck is no longer the calculation. It is the checking. Checking what Claude did in a week took the eye of the field's leading specialist. And what becomes of a proof when nobody can check it any more?
While the researchers slept
We often picture AI as a tool you wield, question after question. This story tells a different one: two researchers write a sentence, switch off the light, and get news every four to six hours.
That morning, the machine was working while they slept. The day that becomes ordinary, the real question will no longer be what it can calculate. It will be who, on our side, still has time to read it through.