# Anthropic Says Claude Computed a Nine-Loop Scattering Amplitude, Beating an Eight-Loop Record

A single prompt, about a week of computation on 96 processors, and a compute cost Anthropic puts at $1,000 to $2,000 per method, verified by the physicist who held the previous result.

- Published: 2026-09-27T06:30:19.445Z
- Canonical: https://polylog.news/ai/2026-09-27/anthropic-says-claude-computed-a-nine-loop-scattering-amplit
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news)

Anthropic published an account on September 25 describing how Claude Fable 5.1 computed, to nine loops, the six-particle scattering amplitude in planar N=4 super Yang-Mills theory (a highly symmetric model physicists use to test calculation methods in quantum field theory). No physicist had reached that level through direct calculation before. The previous result, at eight loops, belonged to Lance Dixon of SLAC National Accelerator Laboratory and Andy Liu, who reached it by an indirect method in 2023. Russian-language AI channels [summarized the claim](https://t.me/ai_machinelearning_big_data/11021) within hours of publication.

The setup deserves close attention. Matt von Hippel, a former theoretical physicist, issued a public challenge on August 7, asking labs to solve a real open problem on the budget of a typical university lab. Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, gave the model a single instruction and then let it work inside Claude Science, checking in periodically. Anthropic reports about a week of computation on 96 processors, at a cost of $1,000 to $2,000 for each of the two methods used.

Verification here is stronger than in most lab-published science demonstrations. Dixon checked the result himself, and the write-up carries von Hippel's name. Even so, the result is a lab-published account that has not gone through peer review. It also addresses a single problem in a narrow, highly structured corner of theoretical physics, where symbolic manipulation and established bootstrap methods do much of the work.

## What this means

What is happening here is cheap autonomous compute substituting for scarce expert time on problems that are hard to solve but easy to verify, which is exactly where these models have the clearest advantage: an independent expert can confirm the answer without repeating the work. Anthropic gains a stronger case for Claude Science as a paid research platform, academic groups with small budgets gain a way around waiting for a specialist, and the assumption being tested is that frontier-level theoretical calculation requires a named research group at a national laboratory. The precedent is computer algebra software in the 1990s, which did not replace theorists but moved the boundary of what counted as routine calculation and shifted credit toward whoever framed the problem. The likelier outcome over the next several months is a series of similar single-prompt results in fields with machine-checkable answers, including formal mathematics, combinatorics and particle-physics amplitudes, provided labs keep funding the compute. This reading would be undercut if attempts on problems without a clean verification step produce results that no expert is willing to certify.

## What to watch

- Whether the nine-loop result is written up and submitted for peer review, since that determines whether the physics community treats it as a citable result or as a vendor demonstration.
- Whether independent academic groups reproduce comparable results on Claude Science or a competing platform at similar cost, which would show whether this works beyond two in-house physicists.
- Whether labs attempt the harder half of the original challenge, the seven-loop divergence question in N=8 supergravity, and what they report if the attempt fails.