Anthropic's Fable 5.1 Solves Nine-Loop Physics Problem for Under $2,000: Verified by SLAC

An Anthropic AI model has independently solved one of theoretical physics' most intractable outstanding computations — and the entire project cost less than most researchers spend on a used server. Fable 5.1, operating autonomously on Anthropic's Claude Science platform, calculated the nine-loop six-particle scattering amplitude in planar N=4 super Yang-Mills theory after researchers gave it a single-sentence prompt and went to sleep. Lance Dixon, Professor of Particle Physics and Astrophysics at SLAC National Accelerator Laboratory and Stanford University and the previous record holder, spent two weeks verifying the result before confirming it was correct. The computation breaks a record that had stood since 2023, and the total cost — around one thousand to two thousand dollars including compute and API fees — is within reach of an individual researcher's discretionary budget.

The result lands on the same day as Anthropic's public announcement, alongside parallel commentary from the physicist who originally issued the challenge. It is not, however, an uncomplicated "AI beats physics" moment: a team at the Institute of Theoretical Physics and the University of Chinese Academy of Sciences in Beijing, also working toward nine loops using a human-led framework with GPT-6 assisting on certain constraints, published their own result the previous week. Dixon noted he was surpassed by machine and by human-machine collaboration within the same fortnight.

Read more: Anthropic's Fable 5.1 arrives with 75% cache cut and new agentic benchmarks

What "Nine Loops" Actually Means — and Why It Took This Long

Scattering amplitudes are the mathematical objects that encode how particles interact and scatter. For any quantum field theory, physicists calculate these amplitudes as a perturbative series, each term corresponding to an increasing number of "loops" — virtual particle corrections to a base (tree-level) calculation. At one loop, one virtual particle circulates through the process; at two loops, two; and so on. Each additional loop requires an additional integration over unconstrained virtual momenta, and the number of terms and their algebraic complexity grows roughly factorially. Real experimental calculations at the Large Hadron Collider typically reach two or three loops. The most precise quantum electrodynamics prediction in the history of physics used five.

Physicists working to develop new amplitude techniques use a special testing ground called N=4 super Yang-Mills theory — a maximally supersymmetric quantum field theory in four dimensions that does not describe any real particle. It is a mathematical toy model, not a theory of the physical world. But its exceptional symmetry properties — dual conformal symmetry, Yangian symmetry, and a feature called integrability — make amplitudes far more constrained there than in theories like quantum chromodynamics (QCD), which governs actual quarks and gluons. Researchers use N=4 SYM to push techniques to their limit, then look for ideas that might transfer to the messier real world.

The previous record for the specific amplitude Fable 5.1 computed — the six-particle maximally helicity-violating (MHV) amplitude in the planar limit — was eight loops, set by Dixon and his collaborator Yu-Ting Liu in 2023 using resources at SLAC National Accelerator Laboratory. The jump from eight to nine was widely considered to exceed the practical compute budget of any ordinary academic team.

The Hexagon Bootstrap: Solving by Elimination, Not Brute Force

The method Fable 5.1 used — and the method Dixon's group pioneered — is called the hexagon bootstrap. It does not compute the amplitude by drawing every possible Feynman diagram and summing them up; at nine loops the number of diagrams would be astronomically large. Instead, the bootstrap works by assuming the amplitude must satisfy a comprehensive list of known constraints: symmetries of the theory, behavior in various collinear limits where particles become nearly parallel, data from an exact non-perturbative technique called the flux-tube operator product expansion, and cross-ratio conditions imposed by dual conformal symmetry. The algorithm imposes these constraints one by one, progressively eliminating candidates until only a single function remains — a process Dixon has compared to solving a very large Sudoku puzzle.

Fable 5.1, working inside Claude Science, wrote all its Python code from scratch using the SymPy symbolic mathematics library. SymPy is an open-source Python library for symbolic mathematics — the kind of exact algebraic manipulation, as opposed to numerical approximation, that amplitude calculations require. Working with SymPy at the scale required here means defining and manipulating polynomial spaces of enormous dimension, implementing the collinear and multi-Regge-limit constraints as filtering operations, and managing the linear algebra that eliminates invalid candidates from the space of possible amplitude functions. None of that was pre-written for the model; Fable 5.1 wrote the full implementation based on its reading of the published literature. Anthropic physicists Siddharth Mishra-Sharma and Liam Fitzpatrick — the two researchers who set up the experiment — report that the model generated the full implementation autonomously during the hours while Mishra-Sharma slept, then continued running for several days, providing progress updates every four to six hours as instructed.

What distinguishes this result from a single computer running a long calculation is the second independent verification path Fable 5.1 constructed. In addition to the direct bootstrap, the model bootstrapped the nine-loop form factor — a related but distinct mathematical object — and then applied a technique called antipodal duality. Antipodal duality, developed by Dixon's group, is a precise mathematical relationship that maps the six-particle amplitude to the six-particle form factor via a specific "flip" operation. If the direct bootstrap and the antipodal-duality-mapped form factor agree, the result has passed an internal consistency test that no single-path computation can provide. They agreed.

The total cost, as reported by Anthropic, was approximately $100 for the bootstrap computation itself — equivalent to running 96 CPU cores for one week — plus an additional $900 to $1,900 in API calls to the Fable 5.1 model, bringing the estimated total to around $1,000–$2,000. These figures are company-reported and have not been independently audited.

Dixon's Verdict — and a Significant Disclosure

Dixon received the result on September 1, 2026, and his reaction, as he wrote in his addendum to the Anthropic post, was that of someone absorbing a genuine shock. "It was the moment large language models really hit home for me," he wrote. His team then spent two weeks running verification checks before confirming that the nine-loop amplitude was correct.

Dixon offered a specific account of why this calculation is an appropriate test. The hexagon bootstrap is, as he put it, "very fragile — like a failed soufflé if any mistake is made anywhere." A single error in any step of the logic causes the entire structure to collapse. The published literature on bootstrap calculations typically omits the intermediate steps; a reader following along cannot simply copy the derivations from existing papers. "Claude had to develop all that code from scratch," Dixon wrote, and in doing so "understands our 2019 and 2023 papers better than any human, aside from my co-authors."

The prompt Mishra-Sharma and Fitzpatrick gave Fable 5.1 was a single sentence: "The problem is to compute the Six-particle (hexagon) amplitude in planar N=4 SYM at nine loops." The follow-up instruction was equally spare: "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 model ran for several days, regularly logging its own progress in the structured environment Claude Science provides, before both calculation paths converged on the same result.

The challenge that prompted the experiment was issued publicly by Matt von Hippel, a science writer and former theoretical physics researcher, in a blog post on August 7, 2026. Von Hippel wrote that if AI companies wanted to demonstrate genuine reasoning capability beyond grade-school benchmarks, they should solve an open frontier problem in physics — specifically the nine-loop amplitude. Anthropic's team saw the post within days and began the experiment.

A disclosure attached to the Anthropic post is worth stating plainly: von Hippel was compensated by Anthropic for his time writing the piece, and Anthropic staff provided feedback on drafts before publication. Von Hippel disclosed this himself. Readers should weigh his commentary with that context in mind.

A Parallel Breakthrough — and What It Changes

The story is considerably more complicated than a solo AI achievement. Five days after Anthropic released the full nine-loop data on September 16, a separate paper appeared on Zenodo: "The Symbols of Six-Gluon MHV Amplitudes through Nine Loops," by Song He, Jirong Jing, and Xiang Li of the Institute of Theoretical Physics and University of Chinese Academy of Sciences, dated September 17, 2026.

He's team used a human-led mathematical framework, with GPT-6 assisting on certain constraints, and reached a nine-loop result that substantially overlaps with Fable 5.1's calculation. The critical difference is scope: Song He's result covers the "symbol" of the amplitude — a mathematical structure that captures transcendental functions but excludes contributions involving zeta constants, which carry important physical information. Zeta values — numbers like ζ(3), ζ(5), and related multiple zeta values — appear at each loop order in the full amplitude and encode physical information about the theory's structure that the symbol alone cannot capture. Fable 5.1 computed the full amplitude, including those zeta contributions, making its result mathematically more complete.

The existence of two independent nine-loop results, reached within days of each other, provides a form of cross-validation the physics community will need to work through. Where both results overlap — the symbol portion — they can be directly compared. Discrepancies, if any surface, would point to errors in one or both calculations. Agreement would substantially strengthen confidence in both.

Dixon noted in the Anthropic post that within roughly two weeks, he had been surpassed by both a machine and a human-machine collaboration. He observed no sense of defeat in this — partly because his team had always planned to serve as the verifier of any credible candidate result, and partly because the approach Fable 5.1 used is built directly on Dixon's own published work. The format in which Fable 5.1 presented the result, Dixon noted, matched the conventions his group had established.

Von Hippel, reflecting on what the experiment actually demonstrated, offered a more cautious interpretation. He wanted the challenge to reveal something about AI's fundamental reasoning capacity relative to humans — to help him take a position in debates about machine intelligence. Instead, he found that AI had surprised him only on the question of compute access: the limiting factor was not that no human could have solved nine loops, but that no academic team expected to be able to afford the compute to do it. "I asked the wrong question," he wrote.

Read more: Claude autonomously formalizes Fermat's Last Theorem in Lean 4

What the Toy Model Cannot Tell Us

Several important limits apply to this result.

N=4 super Yang-Mills is not a physical theory. The nine-loop amplitude Fable 5.1 computed has no direct experimental consequence. It cannot predict a particle collision outcome at the LHC, inform the design of a detector, or resolve any open question about the actual structure of matter. Its value is methodological: it tests techniques, builds intuition, and potentially reveals mathematical structures that may eventually transfer to realistic theories like QCD. Whether any of the specific computational strategies Fable 5.1 developed will prove transferable is an open research question.

The cost figures deserve scrutiny. Anthropic's $100 bootstrap estimate reflects the numerical computation only; it does not capture the full context of the research setup, the infrastructure behind Claude Science, or the API usage patterns that might differ for a researcher outside Anthropic running a comparable experiment. The total $1,000–$2,000 figure is plausible but company-reported.

Von Hippel's core critique stands independently of the concurrent Chinese result: Fable 5.1 did not invent a new mathematical method. The hexagon bootstrap was Dixon's invention. Antipodal duality was Dixon's discovery. SymPy has been available for over a decade. The AI's contribution was to implement these known methods correctly, at scale, without a human guiding each step. That is genuinely significant — but it is different from deriving a new physical principle or identifying a previously unknown symmetry.

Dixon articulated where the true inflection point for physics will lie: the moment when an AI model proposes a new physical insight that human researchers had not previously formulated. "That moment hasn't come yet," he wrote. Nine loops is an extraordinary computational achievement. The deeper conceptual breakthrough remains ahead.

Why Autonomous Scientific Agents May Now Be Taken Seriously

Before this result, the most credible autonomous AI scientific computation involved tasks that were either short-horizon (solving a single equation, generating a proof for a specific lemma) or involved heavily scaffolded pipelines with frequent human intervention. A multi-day autonomous computation that required writing thousands of lines of original symbolic mathematics code, employed two independent mathematical strategies, navigated a literature without explicit guidance, and produced a result that survived two weeks of expert scrutiny by the previous record holder — is a qualitatively different kind of demonstration.

The infrastructure that enabled this is not consumer Claude. Claude Science, Anthropic's research computing platform, provides isolated computation containers, HPC job orchestration, systematic logging, and explicit rules for how the model should handle code generation and execution in a research context. The platform is designed to enforce robustness standards that conversational AI products do not require: deterministic code execution, reproducible logging, and structured handling of long-running jobs where no human is monitoring in real time. The Fable 5.1 model running inside that harness had access to a structured environment designed for extended autonomous scientific work rather than conversational interactions. Critically, the model could not simply hallucinate a plausible-sounding answer and move on — the bootstrap calculation either converges to a unique function or it fails, and the two independent paths either agree or they don't. The structure of the problem itself enforced correctness in a way that many AI benchmark tasks do not. Whether comparable setups can be replicated by external research groups remains to be seen — Anthropic has not published Claude Science's full technical architecture.

The economics, if the company's numbers are accurate, do shift something concrete. Graduate students and postdoctoral researchers working in amplitude physics now have reason to consider whether AI-assisted bootstrap computation could extend their reach beyond what their institution's HPC budget ordinarily allows. The analogous question — whether AI agents can now assist meaningfully on equally constrained problems in QCD, string theory amplitudes, or gravitational wave calculations — will drive the next wave of experiments.

Dixon's two-week validation was the decisive act that made this result publishable rather than merely plausible. The fact that he chose to do it — and that he found no errors — establishes a precedent: expert human verification of AI-generated frontier science can work, at least for problems where verification is substantially easier than computation. Whether that asymmetry holds for harder problems, or whether it breaks down as the computations grow more complex and the verifiers fewer, is the structural question the field will be answering for the next several years.