Claude sets a physics record: what happened and why it matters

What happened: Anthropic said on 25 September that its Claude AI model computed the nine-loop six-particle scattering amplitude in a physics theory called planar N=4 super-Yang-Mills, beating the previous eight-loop record set in 2023 by physicists Lance Dixon and Andy Liu.

The background: Scattering amplitudes predict how particles behave; extra "loops" add precision but make the maths far harder. The theory used, N=4 super-Yang-Mills, is a simplified physics testbed, not a model of real particles.

How it happened: Claude answered a public challenge posted on 7 August by physics writer Matt von Hippel. Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma gave the model, running as Fable 5.1 inside Claude Science, one instruction and let it work largely alone from there.

The effort: Claude completed the calculation two separate ways, taking about a week on 96 CPUs, at a cost of roughly $1,000 to $2,000 per method, Anthropic said.

The validation: SLAC physicist Lance Dixon spent about two weeks independently checking the result. A separate Beijing team led by Song He reached most of the same result around the same time, using GPT-6-based AI assistance for some constraints.

The endorsement: Dixon praised the model's ability to carry out every step of the complex calculation method involved, calling the achievement a genuine triumph for a large language model.

Why it matters: the result adds to a run of high-profile claims from AI labs about models tackling frontier science problems, and doubles as validation for Claude Science, Anthropic's push into research tools.

The caveat: Anthropic was clear this isn't new physics — Claude executed a known calculation recipe on an unusually hard problem, rather than discovering something new about how particles behave.

The scale: the previous eight-loop record, set by Lance Dixon and Andy Liu in 2023, had stood for about three years before Claude's nine-loop result surpassed it.

Independent check: a Beijing-based team's parallel effort, reaching a similar answer using different AI assistance, adds further confidence that the nine-loop result itself is correct.

What's next: Anthropic did not say whether it plans a ten-loop attempt, describing the result instead as a demonstration of Claude working largely unsupervised on a hard, well-defined task.

The cost: Anthropic put the bill at roughly $1,000 to $2,000 per method, for about a week of computing on 96 CPUs, a small outlay for a record that no human team had reached by direct calculation.

The origin story: von Hippel put the nine-loop problem to AI companies in general as a public challenge, and Anthropic's own team chose to take it on.

The trust angle: an independently checked result, largely matched by a separate Beijing-based team, is harder to dismiss than a typical internal AI benchmark score alone.