Anthropic's Claude Ran a Week on 96 CPUs and Computed a Nine-Loop Physics Amplitude That Stood Out of Reach for 12 Years — for About $2,000

Anthropic published a blog post on September 25, "Yes, Claude can do Nine Loops," written by theoretical physicist and science writer Matt von Hippel, reporting that its Claude model completed a frontier calculation in quantum field theory with almost no human supervision. Chinese science media amplified the result on September 29, and the detail drawing the most attention is the invoice: roughly \$1,000 to \$2,000 in total, of which the core numerical computation cost about \$100.

The problem is the six-particle scattering amplitude in planar N=4 super-Yang-Mills theory — the simplified model physicists use as a testing ground for new methods. Scattering amplitudes predict how particles behave in collisions, and they are computed in "loops," where each additional loop brings the answer closer to exact but multiplies the complexity. Most amplitude calculations in physics stop at two or three loops; Dixon's team at SLAC National Accelerator Laboratory pushed from three loops in 2011 to eight loops in 2023 — five loops in twelve years — and nine loops was the next distant summit.

Claude was given a single line of instruction: compute the nine-loop amplitude, then a homely note that the human was going to sleep and would check in every four to six hours. Running unsupervised, the model wrote complete Python scripts using the SymPy symbolic math library from scratch, implementing the "bootstrap" method the field had developed — constraining the space of candidate functions with physical principles until only one answer survives. The heavy numerical part ran on 96 ordinary CPU cores for about a week, at a power-and-server cost of around \$100, according to figures Anthropic reported.

The result came with its own redundancy check. Claude ran two independent routes — the direct bootstrap, which Dixon himself had considered too hard for nine loops, and the indirect route through form factors that Dixon's team used for its 2023 record. Both routes agreed on every one of the 107,053 nonzero coefficients. Lance Dixon then spent two weeks verifying the code and result with his team's own tools before endorsing it; Anthropic's physicists Liam Fitzpatrick and Siddharth Mishra-Sharma co-led the project.

There is a twist in the race: a Beijing group led by Song He, with Jirong Jing and Xiang Li, posted the nine-loop symbols on Zenodo days before Anthropic's writeup, having reached the same milestone independently the same week. The human race and the machine run converged on the same frontier within days of each other.

Caveats matter, and Anthropic's own post is careful about them. Claude discovered no new physics — it executed methods published by human researchers in 2019 and 2023, and, as Dixon noted, the method details still await independent scrutiny and digestion by the physics community at large. What the result demonstrates is reliability: an LLM carried a fragile, error-prone, multi-day symbolic computation to completion without a scientist watching each step.

The significance is less the loop count than the cost curve. A calculation that a top human team treated as a multi-year goal was executed for the price of a decent laptop, on hardware any university lab already owns. Physics is now the third discipline — after mathematics and biology — where AI systems have completed original research-grade computations inside human frameworks. The tools have not started asking new questions yet. But the era of "AI as a research instrument," rather than a demo, has arguably begun.