Skip to content
NewsResearch

Anthropic physicists used Claude to push a 2023 physics record from eight loops to nine

· by Pondero Newsdesk

The short version

Two Anthropic physicists used Claude to compute a nine-loop scattering amplitude in planar N=4 super Yang-Mills theory, one loop past the 2023 record, in a paid guest post that disclosed financial ties to the researcher who validated it.

Anthropic physicists used Claude to push a 2023 physics record from eight loops to nine

A physics calculation that had stalled at eight loops since 2023 got a ninth loop from an unsupervised overnight run of Claude. Science writer Matt von Hippel disclosed the result in a paid guest post he published on Anthropic's research site on September 25, 2026, and the method that solved it cost about $100 in compute, per the post.

What

Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, used the company's Fable 5.1 model inside Claude Science, a paid research harness, to compute the six-particle scattering amplitude in planar N=4 super Yang-Mills theory at nine loops, one loop past the record Stanford physicist Lance Dixon and collaborator Andy Liu reached indirectly in 2023, according to the guest post. The work answered a public challenge von Hippel posted to his blog on August 7, 2026, asking AI companies to solve a computationally hard amplitudes problem using only the compute budget an academic could access. Fitzpatrick and Mishra-Sharma gave Claude a one-line prompt describing the task, then let it run unattended overnight with instructions to report back every four to six hours. It returned two independent derivations that agreed on all 107,053 nonzero coefficients in the harder of the two result files, according to Unite.AI's reporting on the post. The bootstrap method, written in Python with the SymPy library, ran on 96 CPUs for a week and cost about $100 of the total; either full approach would cost an end user "around one or two thousand dollars," the post states.

The guest post disclosed a paid relationship, not just a research result

Anthropic's post disclosed that it invited von Hippel to write the piece and compensated him for his time, and that Dixon, who spent two weeks independently validating the nine-loop result, received Claude usage credits for that work, per Unite.AI's account of the post's disclosure. Von Hippel wrote that Claude executed the calculation with no scientific oversight beyond repeated instructions to keep going, using known bootstrap methods rather than a new technique.

A rival team reached the same result days later without Claude

Song He, an amplitudeologist at the Chinese Academy of Sciences, told von Hippel his group had independently obtained most of the same nine-loop result using GPT-6 assistance for some of the constraint-checking rather than a single unsupervised run. He and two collaborators at the Institute of Theoretical Physics and the University of Chinese Academy of Sciences published their own dataset on Zenodo on September 17, 2026, under a Creative Commons license, per Unite.AI. Von Hippel, who set the original challenge, wrote that the outcome mostly showed there is more low-hanging fruit in the field than he expected, not the qualitatively new capability that would settle debates about superintelligence.

The compute is cheap to budget; the validation needs to be independent of the vendor

For a research or ops team pricing out an unattended, multi-day agent run, this result sets a real budget line: $100 to $2,000 in compute, per the guest post, bought a calculation three years of expert effort had not reached. Do not price the validation the same way. The physicist who checked the answer was paid in Claude usage credits, and the writer who reported the result was paid by Anthropic for the piece, per Unite.AI's account of the disclosure. Treat the compute figure as real and budgetable, and route the output to a reviewer with no financial tie to the vendor before a team relies on it for its own research.

Sources