Anthropic physicists ran a largely unsupervised calculation on a paid science harness; SLAC’s Lance Dixon checked the answer. Cost sat in the low thousands.
A former amplitudes researcher put a public dare on the table in August. A month later, two physicists at Anthropic said Claude had met one of the targets.
Matt von Hippel, now a science writer who posts weekly at 4gravitons.com, used to compute scattering amplitudes. Those formulas tell physicists how likely particles are to interact given their momenta and energies. Most such formulas stop at two loops. A few reach three. The electron’s anomalous magnetic moment is famous for five. In a toy theory — planar N=4 super Yang-Mills — the record for a particular six-particle amplitude sat at eight loops.
Von Hippel wanted a test that felt computationally hard, not merely clever. He wanted something an academic group might have skipped because the machines and the weeks looked too expensive. On Aug. 7 he wrote:
“If AI companies want to impress people like me (or scare us, for that matter), then they need to tackle my old field. Show that an AI can take the kinds of computer resources an academic has access to, and solve one of the scattering amplitudes field’s big outstanding problems. Show that a computational limit everyone expected to be a problem doesn’t actually matter. Give us N=8 supergravity to seven loops, or N=4 super Yang-Mills to nine loops.”
N=4 super Yang-Mills is not a model of the real world. Three of the four forces are Yang-Mills theories. The “N=4 super” part piles on four supersymmetric partners per particle. That surplus makes the theory unrealistic and, paradoxically, easier to calculate. Amplitudeologists use it to stress-test methods.
Von Hippel helped compute a three-loop amplitude in graduate school and later watched the field reach seven. Lance Dixon, a professor at the SLAC National Accelerator Laboratory, and collaborators pushed the same six-particle object to eight loops in 2023, in part through a related form factor and a symmetry called antipodal duality.
At the end of August, Liam Fitzpatrick and Siddharth Mishra-Sharma, physicists at Anthropic, contacted von Hippel. After Dixon checked the output, they walked through the run.
They used Fable 5.1 inside Claude Science, a paid harness that wraps Claude with structured rules. After asking which problem looked most tractable, they gave a short prompt:
“The problem is to compute the Six-particle (hexagon) amplitude in planar N=4 SYM at nine loops.”
Then they kept it moving. One instruction read:
“I’m going to sleep and won’t be available for another several hours. Keep working on this until I tell you to stop. Give me updates every 4-6 hours.”
Claude produced the result two ways: the original bootstrap, and the indirect form-factor route. Either path would have cost an end user about $1,000 to $2,000, mostly from running the model so long. The bootstrap, in Python with SymPy, accounted for about $100 of that — 96 CPUs for a week.
A few days later, Song He of the Chinese Academy of Sciences said his group already had most of the result. They used some GPT-6 help. Not the same nearly hands-off setup.
Dixon, in an addendum dated with the Sept. 25, 2026, guest post, said the news landed on Sept. 1. “For me, it happened on September 1, when Liam Fitzpatrick and Siddharth Mishra-Sharma at Anthropic told me that Claude had computed the nine-loop MHV six-particle amplitude in planar N=4 super Yang-Mills, and asked me to validate its result.”
He had expected the direct amplitude to be too brittle. “I thought it would be too hard to do the amplitude directly. So I was really quite impressed that Claude could do it directly.” A single mistake in the recipe, he wrote, and “it all crashes down like a failed soufflé.”
Does being scooped by a machine bother him? “No, for two reasons.” His group already had a campaign to validate machine-generated candidates. And Claude used the methods Dixon’s collaborators built. “In fact, I would assert that Claude understands our 2019 and 2023 papers better than any human, aside from my co-authors.”
The humans — Dixon, He, and collaborators — will publish and analyze the formulas. Computer-readable files matching earlier six-, seven-, and eight-loop formats were posted.
Von Hippel’s read is narrower than a new physical principle. Claude used known methods and more compute than people had casually thrown at one extra loop. His largest takeaway: experts can misjudge how reachable a well-defined goal is. Computer scientists who said amplitudeologists needed programmers, he wrote, should feel vindicated. The harness finished without an outside collaborator steering each step.
He compared the work with March projects that still looked like student exercises. This one sits at the expert frontier of a small subfield. How far that generalizes to real-world amplitudes, he said he does not know. Groups already use AI for coding. If they have not tested whether a science harness can one-shot another loop there, he argued they should — and they should know how they will check the files.
Anthropic invited von Hippel to write the post and paid him for his time. Staff gave draft feedback. The opinions are his. Dixon validated independently and received Claude usage credits.