Claude’s Nine-Loop Physics Calculation: What It Shows

Claude reached nine loops on a calculation that had previously topped out at eight. Anthropic researchers Liam Fitzpatrick and Siddharth Mishra-Sharma used the Fable 5.1 model in Claude Science to calculate a six-particle, or hexagon, amplitude in planar N=4 super Yang-Mills, a theoretical model rather than a direct prediction for an observed system. The reported result is a higher-order calculation, not a new theory of nature.

A narrow milestone, with a demanding workflow

The researchers started with a short prompt asking for the nine-loop amplitude. They then asked Claude to continue and provide updates. The model produced the result using both bootstrap and form-factor approaches, established methods used in this area to extend calculations to higher loop orders. The article estimates the compute cost at thousands of dollars. Those details make the achievement more concrete than a general claim that an AI can do physics: it describes a specific task, methods, and level of human direction.

The result builds on a long-running effort. For this six-vertex calculation, progress moved from three loops to eight between 2011 and 2023. The nine-loop calculation extends that frontier by one order. The earlier results and new loop order do not mean Claude independently invented the underlying framework: the article says the work relied heavily on existing research, and other groups had mostly already solved the problem.

Verification mattered as much as the answer

Lance Dixon, who had developed a framework for checking a nine-loop solution, verified the result after receiving it on September 1, 2026. The calculation used two established routes, but agreement across approaches and expert review are especially relevant because the construction is fragile: the article notes that a mistake in the computational recipe could cause it to fail. Dixon’s verification and the reported fragility put the model’s role in context. Producing a candidate answer is not the same as establishing that it is correct.

The article also reports that Claude had to develop code for the construction. That is a meaningful form of technical execution, but it does not establish broad autonomy or a new research method. The accomplishment lies in carrying out a difficult calculation using known approaches, with limited prompting and a human expert able to check the result. The reported workflow and its limits support that narrower claim.

What the result does, and does not, establish

Higher-loop calculations add contributions that can make quantum-field-theory predictions more precise. But this one concerns planar N=4 super Yang-Mills, a finite, scale-invariant toy model in the planar limit of a four-dimensional theory. It is interesting to theorists, including because of its place in discussions of string theory, but the nine-loop result is not itself a new experimental finding or a calculation for a directly observed physical system. The model’s scope is central to judging the achievement.

The useful signal is specific: Claude helped push a technically delicate calculation beyond the earlier reported frontier, while relying on prior work and expert verification. That is a stronger demonstration of research workflow capability than a fluent explanation would be, but it is not evidence that the model can independently choose and solve open problems across theoretical physics.