Claude computes a nine-loop scattering amplitude, breaking a three-year-old physics record

Anthropic has published a guest post by physicist Matt von Hippel, “Yes, Claude can do nine loops,” reporting that Claude computed the nine-loop six-particle scattering amplitude in planar N=4 super-Yang-Mills theory. The previous record, eight loops, was set in 2023 by SLAC’s Lance Dixon and collaborators.

The facts

• The challenge: on August 7, von Hippel dared AI companies on his blog to solve an open scattering-amplitude problem using compute an academic could afford. Nine loops was one of his two options.
• How it ran: per Jiqizhixin’s report, two Anthropic physicists used Fable 5.1 on the Claude Science platform. The initial task was a single sentence, and later guidance was mostly “keep going.” Claude computed the result twice, once with a direct bootstrap and once via Dixon’s indirect form-factor route.
• Cost: roughly $1,000–2,000 at consumer prices, mostly for running the model over days. The numerical bootstrap in Python and SymPy cost about $100, equivalent to 96 CPUs for a week.
• Verification: Dixon spent two weeks independently checking the result. Separately, Song He’s group at the Chinese Academy of Sciences’ Institute of Theoretical Physics published a symbol dataset covering two through nine loops on September 17, arriving at nearly the same time.

Our take

The notable part is long, low-supervision execution. The model had to build a many-step pipeline where any single mistake collapses the whole result, then keep it running for days. That’s the same capability long-horizon coding agents depend on.

Don’t overread it, though. The nine-loop result used methods Dixon’s team developed over years; Claude carried an existing recipe all the way through rather than proposing new physics. Dixon himself says the unsettling moment will come when a model proposes new physical principles first.