Claude’s Nine-Loop Physics Result Explained: What It Solved, What It Cost, and Why It Matters

Claude has completed a difficult calculation at the research frontier of theoretical particle physics. The achievement deserves attention because of the work it sustained: turning established methods into working code, managing a long computation, and delivering a result that a specialist could check.

The scope matters. This is a calculation in a simplified theory used to study particle interactions. It establishes neither a new law of nature nor a general ability to replace physicists. The public research files also distinguish between an extensively cross-checked mathematical representation called the symbol and a fuller function whose independent verification is less complete.

Here is what the announcement says, what the released files contain, and what researchers can reasonably take from the result.

Evidence checked September 25, 2026. Kingy inspected the linked research documentation and performed the limited file and arithmetic checks described below. We did not reproduce the nine-loop calculation or independently validate its physics.

What did Claude calculate?

The target was the nine-loop, six-gluon, maximally helicity-violating amplitude in planar N=4 super-Yang–Mills theory. Readers do not need to unpack every term to understand the achievement: it is a precisely defined mathematical description of a particular particle-scattering process in a highly symmetric model.

The earlier eight-loop result was published by Lance Dixon and Yu-Ting Liu in 2023. Their method connected the amplitude to a related quantity called a form factor through a relationship known as antipodal duality. That work provided an important foundation for the new calculation. Dixon and Liu’s eight-loop paper.

Anthropic’s September 25 guest post, written by physicist and science writer Matt von Hippel, reports that Fable 5.1 tackled the next loop order through Claude Science. Dixon, a professor at SLAC and Stanford, describes checking the result in an accompanying account. Announcement and Dixon’s account.

What does “nine loops” mean?

A scattering amplitude helps physicists calculate the probabilities of different outcomes when particles interact. These calculations often use a perturbative expansion: start with a leading contribution, then add successive orders of quantum corrections. Loop order labels those corrections. It has nothing to do with the number of times an AI agent retries a task.

Higher orders can improve a prediction within the expansion’s useful regime, but the calculations become increasingly demanding. Moving from eight to nine loops does not mean a fixed percentage improvement in accuracy, and it is not comparable to adding one more ordinary spreadsheet row.

The theory here is a deliberately simplified testing ground. Its exceptional symmetry makes calculations tractable enough to investigate methods that would be much harder to develop directly for real-world particle processes. A successful calculation in this setting is valuable theoretical work; applying the same approach to less symmetric theories remains a further challenge.

How Claude approached the problem

The announcement describes a short initial task followed by instructions to continue working. Claude Science supplied the environment around the model, allowing it to pursue a computation over an extended period. Anthropic reports two approaches: a direct amplitude bootstrap and a route through the related form factor. Research account.

“Bootstrap” means constraining an answer using properties it must satisfy. Researchers choose a mathematical space that could contain the answer, then impose symmetry, consistency conditions, and behavior in known limits. Enough constraints can select a unique candidate within that space. The choice of space and the validity of the constraints still matter.

That dependence on previous research is central to the story. The form-factor route builds on work by Dixon, Omer Gürdoğan, Andrew McLeod, and Matthias Wilhelm, as well as the later eight-loop amplitude calculation. Form-factor paper.

Our reading is that the strongest demonstrated capability is sustained execution of a demanding research workflow. The public material gives much less reason to claim that Claude invented a new physical principle. It shows how useful reliable implementation can become when experts have already developed powerful methods.

What did it cost?

Von Hippel reports an estimated end-user cost of roughly $1,000–$2,000 for either approach, with most of the expense attributed to running Claude. The direct bootstrap’s computation reportedly accounted for about $100, corresponding to 96 CPUs running for a week. These are the source’s estimates, not a Kingy invoice or a current price quote. Reported costs.

Those figures describe different parts of the work:

A full research budget would also account for expert review, setup, and the earlier work that made the method possible. For a lab assessing AI assistance, the useful measure is the cost of an output that survives a defined check. Cheap computation has limited value if nobody can establish what it produced.

What the published outputs actually establish

The Cosmic9 research page supplies computer-readable results, sample coefficients, conventions, checksums, and validation records. Its documentation is more specific than the social announcement.

The crucial distinction is between a symbol, which retains important mathematical structure while omitting certain information, and the full function, which requires additional terms and constants. Checking the symbol does not automatically check everything needed for the function.

The method and validation note also identifies checks that remain undone. Some final ambiguities depend on particular flux-tube inputs without an independent check sensitive to them. The assumed mathematical space may omit possibilities outside it. The computation programs themselves are not distributed, which limits straightforward end-to-end reproduction. The function-specific qualification appears on the result page.

Dixon’s expert validation is significant evidence. These documented boundaries specify how far readers should extend it. A dataset release, an expert check, and a fully independent rerun are different forms of support.

A concurrent human-led result deserves credit

Song He, Jirong Jing, and Xiang Li released The Symbols of Six-Gluon MHV Amplitudes through Nine Loops, dated September 17, 2026, on Zenodo. Their archive contains symbol data through nine loops and a Wolfram Language script for accessing coefficients. Its README explicitly excludes the information invisible to the symbol, including zeta constants. Concurrent dataset.

The Anthropic account says this group used GPT-6 assistance for some constraints, while humans handled the overall framework. That attribution comes from the account; the deposit itself establishes the authorship, scope, date, and available files. It would be misleading to turn the story into a contest in which human researchers had no route forward.

What Kingy checked

We downloaded four selected Cosmic9 files and confirmed that their SHA-256 hashes matched the published manifest. We also checked the downloaded concurrent archive against Zenodo’s checksum, inspected its file listing, and read its README.

For a narrow arithmetic check, we took the rational coefficients listed in Cosmic9’s 20,630-row sample and reduced them modulo the two stated primes. Every result matched its listed residue. We repeated that operation for the exact coefficients in all 107,053 rows of the published comparison record; all matched both listed residues. This checks consistency inside the supplied records. It does not recompute those coefficients from the underlying amplitude data. Sample file, comparison record, and checksum manifest.

We did not run either research campaign, execute the concurrent group’s Wolfram script, or establish equality between the two groups’ datasets. Our checks support the integrity and internal arithmetic consistency of selected published materials. Scientific validation remains attributed to the researchers.

Practical lessons for AI-assisted research

This case suggests three useful habits.

Specify an answer someone can check. A defined calculation creates a clearer target than an open-ended request for a discovery. For an everyday project, that could mean a reproducible analysis with named inputs, a saved script, and expected outputs.

Budget for persistence and review. A multi-day agent run needs a spending limit, progress records, and a stopping condition. Include the time required to inspect the result. The reported cost here is evidence about one task, not a universal price for scientific progress.

Match the claim to the verification. File checksums, internal consistency, expert review, and independent reproduction answer different questions. Keep those distinctions in the final report. For a broader framework, see Can AI Make Scientific Discoveries? Real Examples and What Counts as Proof.

If you want a manageable starting task, our practical guide to Claude for scientific research includes reusable prompts and a public-data analysis. Begin with an output whose correctness you can inspect before attempting a frontier calculation.

Disclosures and source limits

Anthropic commissioned von Hippel’s guest post and compensated him for his time. Its staff commented on drafts; the post says the opinions are his own. Dixon validated the result independently and received Claude usage credits. These are the source’s disclosed relationships. Original disclosure.

This explainer draws on that account, the released data and method notes, the concurrent Zenodo deposit, and the earlier research papers. Kingy’s original checks are limited to those explicitly described above. The next substantial evidence to watch for is broader independent checking of the full function, together with enough computational detail for other teams to reproduce the result.