Claude computes a nine-loop amplitude in N=4 super-Yang-Mills

A guest post by the physicist and science writer Matt von Hippel, with an addendum by Lance Dixon, on how two Anthropic physicists used Claude in the Claude Science harness to compute the nine-loop six-gluon amplitude in planar N=4 super-Yang-Mills that von Hippel had set as a challenge; Dixon validated the result, a human group with AI assistance reached the symbol at the same time, and von Hippel concludes the problem was less out of reach than experts had thought.

Matt von Hippel (guest post), with an addendum by Lance Dixon (SLAC and Stanford) · 2026-09-25 · original

On 7 August 2026 the physicist and science writer Matt von Hippel posted a challenge on his blog, 4gravitons: an AI company that wanted to impress amplitude physicists should take "the kinds of computer resources an academic has access to" and solve one of the field's outstanding computational problems, such as "N=4 super Yang-Mills to nine loops." This guest post, published on 25 September 2026 under Anthropic's Science label (its page heading is "Yes, Claude can do Nine Loops"), is his account of what happened next: two Anthropic physicists had Claude compute the amplitude within a month. It ends with an addendum by Lance Dixon, the physicist who validated the result. The post is a first-person account of one project, not a study. The version of record for the result itself is the data release linked from the post (see "Who checked the result" below); no paper describing it had appeared on arXiv as of 13 November 2026.

What the problem is

Scattering amplitudes and loops

The post explains the problem for non-specialists (section "The challenge"). Scattering amplitudes are formulas that give how likely subatomic particles are to interact in particular ways. They are computed as a series of approximations ordered by loops, a measure of how complicated the allowed interactions are; each loop brings the result closer to the full answer and makes the calculation harder. Most amplitudes in particle physics are known to two loops, a few to three.

N=4 super-Yang-Mills is a "toy model": its extra supersymmetric particles make it unrealistic but easier to compute in, so it is used to develop and test methods. The target was the six-particle (hexagon) MHV amplitude in the planar limit of the theory, which Dixon and collaborators had computed through eight loops.

The bootstrap and the form-factor route

These results were obtained with a bootstrap: rather than summing particle interactions, one writes down the space of functions the answer could be, in a specialized alphabet, and then removes possibilities using known constraints (limits, symmetries, results from other methods) until one candidate is left with checks to spare. Von Hippel compares it to Sudoku. Dixon's eight-loop result (Dixon and Yu-Ting "Andy" Liu, arXiv:2308.08199) took an indirect route through a related, easier object, a form factor, using a symmetry called antipodal duality to carry the form factor over to the amplitude. Dixon writes in the addendum that since 2023 his group had aimed to reach nine loops by this route, and that he "thought it would be too hard to do the amplitude directly."

What the post reports

How the calculation was run

According to von Hippel (section "Then people did it"), Liam Fitzpatrick and Siddharth Mishra-Sharma, physicists at Anthropic, contacted him at the end of August to say they had met the N=4 challenge. They used the model Fable 5.1 inside Claude Science, which the post describes as a commercial harness that runs Claude "with structured rules and prompts." After asking Claude "which problem it was most likely to be able to tackle," they gave it a one-line task, quoted in the post, and then mostly told it to continue (see "The prompts" below).

Claude did the calculation two ways, by the bootstrap and by the form-factor route. The post gives the costs as estimates: either approach "would have cost an end-user around one or two thousand dollars," mostly for running Claude, and the bootstrap's own computation, written in Python with the SymPy package, used about $100 of that, "corresponding to running 96 CPUs for a week." The post does not say how long the whole project took in wall-clock time, how many sessions it used, or how many attempts failed along the way; von Hippel writes that he does not know "how many mistakes Claude made internally."

A concurrent human result

A few days after Anthropic's contact, Song He of the Chinese Academy of Sciences in Beijing reported that his group had already obtained most of the result, with AI assistance based on GPT-6 for "some of the constraints, but not for the overall framework" (von Hippel's section and Dixon's addendum). The post links their data release: He, Jing and Li, The Symbols of Six-Gluon MHV Amplitudes through Nine Loops, deposited on Zenodo on 17 September 2026, which contains the symbol (the amplitude's leading structure, which leaves out terms carrying transcendental constants such as zeta values) from two through nine loops. The post does not say whether the two nine-loop symbols were compared or agree, and the Anthropic data page does not mention the He group's result.

Who checked the result

Dixon's validation. The addendum is by Lance Dixon, Professor of Particle Physics and Astrophysics at SLAC National Accelerator Laboratory and Stanford University. Fitzpatrick and Mishra-Sharma told him of the result on 1 September and asked him to validate it. Because the nine-loop form factor can be recovered from the amplitude, he validated "mostly that way," over about two weeks. His group had stated in advance that it had "all the tools to validate any candidate solution a machine would provide us." The post's disclosure says he "validated the result independently and received Claude usage credits."

The data release. The post links the full result on a page hosted by Mishra-Sharma (smsharma.io, "Cosmic9"), laid out like Dixon's page for the lower loop orders and dated 16 September 2026. It gives more detail than the post about how the result was obtained and how far it is certified:

  • The nine-loop symbol was obtained by the form-factor route: the nine-loop form factor was bootstrapped, mapped to the amplitude by antipodal duality on a surface where that applies, and then extended off that surface, with remaining ambiguities fixed by the amplitude's behavior at the origin and by flux-tube (near-collinear) data. The method note lists both among its assumptions, with antipodal duality at nine loops and the persistence of the known alphabet (validation/method_and_validation.md, section 5).
  • The calculation was done in modular arithmetic. The form factor is certified over the rationals at three 31-bit primes; the amplitude was computed at two primes. Of the 1,018,297 nonzero coordinates of the main representation, 1,014,476 (99.62%) reconstruct to certified rational numbers and 3,821 do not.
  • A separate direct bootstrap of the nine-loop symbol produced a second representation, and the page reports that the two "agree on every coefficient compared," all 107,053 nonzero coefficients that determine it, 3,401 of them compared modulo the two primes only.
  • As a control, the same programs run at eight loops agreed with the published eight-loop symbol on 1,000 random words, checked modulo one prime.
  • The method note's section 8, "What remains untested," adds that every check is at symbol level, that five directions fixed last by the flux-tube data have no independent check, and that the multi-Regge limit was not tested quantitatively at nine loops.
  • The amplitude as a full function, beyond the symbol, "has been computed once, and there is no second, independent computation of it," and rests on one further stated assumption.
  • "The programs of the computation are not distributed."

So the outside check on the result is Dixon's validation; the further checks listed on the data page compare Claude's own computations with each other and with published lower-loop results. The post says that the humans involved, Dixon, He and their collaborators, "will get to publish the results." No peer-reviewed paper or arXiv preprint on the nine-loop amplitude was found as of 13 November 2026 (arXiv search; Dixon's most recent listed papers do not include one).

What the author concludes

Von Hippel's conclusions (section "So, problem solved?") are more measured than the post's title:

  • He had hoped to see AI overcome a computational barrier in a surprising way. Instead, he writes, "it did something it turned out humans were also able to do": Claude used known methods with somewhat more computing than people had tried, perhaps helped by using Python rather than Maple or Mathematica and by better software engineering, "but not super-intelligently so."
  • His "biggest takeaway" is that "there is more low-hanging fruit out there than you'd expect": a well-defined goal can look less achievable to experts than it is.
  • He calls it noteworthy that the harness reached the end without scientific oversight beyond "keep going," and addresses readers who think AI too error-prone to use: "It can do this kind of thing reliably now." The note observes that this general claim rests on one calculation.
  • He contrasts it with the student-level work of March 2026 (the Vibe physics post) and judges that "the technology has genuinely gotten better," while allowing that this may be an unusually AI-friendly problem.
  • He is unsure how far it generalizes. Toy-model amplitudes are small sub-communities; real-world amplitude calculations are more competitive and may have less low-hanging fruit.
  • On his original question, whether AI can get around computational limits in ways relevant to debates about superintelligence, he writes that he "didn't really get an answer."

Dixon's addendum calls the computation "quite a triumph" for a language model, notes that Claude used his group's methods and format, and offers the opinion that "Claude understands our 2019 and 2023 papers better than any human, aside from my co-authors." The "more soul-searching moments," he writes, will come when models produce new physical principles before humans do.

The prompts

The post quotes two prompts. The task was stated in one line: "The problem is to compute the Six-particle (hexagon) amplitude in planar N=4 SYM at nine loops." Later instructions were of the form "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." The post does not publish the Claude Science harness's own rules and prompts, which it describes only as structured, and it does not compare alternative wordings.

Limitations and open questions

  • One case, chosen problem. The account covers one calculation. The problem was chosen after asking Claude which problem it was most likely to be able to tackle; the post does not report attempts on the other (N=8 supergravity at seven loops).
  • Disclosed interests. Anthropic invited and paid von Hippel; Anthropic staff commented on drafts, and the post states that the content and opinions are his. Dixon received Claude usage credits. In a companion post on his blog the same day ("It Got to My Field," 4gravitons, 25 September 2026), von Hippel says he signed no agreement with Anthropic, that it proposed only light edits, and that he has not seen "the logs from the LLM," so his account of the run rests on what the two physicists told him. He also writes there that the team solved the problem first and contacted Dixon afterwards, and that "90% or more of the compute cost was from the LLM, not the calculation itself."
  • Verification so far. The result was checked by one outside expert and by the computation's own internal cross-checks. The full function has one computation; 3,821 of the main symbol representation's 1,018,297 nonzero coordinates, about 0.4% (our arithmetic), did not reconstruct to certified rational numbers and are known only modulo the primes; the programs are not released, so the computation cannot be rerun by others. No publication had appeared by 13 November 2026.
  • Costs are estimates. The dollar figures are what the computation "would have cost an end-user," not reported invoices, and human time spent by the two Anthropic physicists is not given.
  • Not a new method. Both von Hippel and Dixon say Claude used existing methods. The novelty claimed is that the calculation was carried out, with little supervision, at modest cost.

Research questions

  1. Do the Anthropic and He group nine-loop symbols agree? Two independent results now exist. A coefficient-by-coefficient comparison of the two public releases, in a common normalization, would settle it and would be the strongest outside check available.
  2. How much expert supervision did the run actually involve? The post reports prompts like "keep going." Publishing the session logs, or a count of human messages and their content, would show whether any physics guidance was given beyond the task statement.
  3. Was the N=8 supergravity challenge attempted? Claude was asked which problem it was more likely to solve. Reporting whether the other challenge was tried, and how it went, would show whether the success depends on problem selection.
  4. How sensitive is the outcome to the harness? The same model with and without the Claude Science harness's rules, on the same task, would separate the contribution of the harness from that of the model.
  5. Can the remaining 3,821 coordinates be certified? Computing at further primes, as the separate direct bootstrap did, would show whether the uncertified part of the representation hides any inconsistency.
  6. Does the approach transfer to real-world amplitudes? Von Hippel suggests that groups working on real-world amplitudes test whether science harnesses can "one-shot frontier calculations," with a plan for checking the results. A documented attempt on a calculation with a known human result would test his point about low-hanging fruit in a more competitive area.

How it connects

References