Anthropic says Claude Fable 5.1 completed a nine-loop six-gluon amplitude calculation in planar N=4 super-Yang–Mills, extending a difficult human-developed computational workflow. The result matters because it is a credible example of an agent carrying a brittle expert process through code construction, debugging, symbolic algebra, and long-running compute—not because it discovered a new law of physics. Its strongest caveat is equally clear: the full calculation was not independently rerun end to end with separate released code.
Key facts
- The target was a six-particle MHV amplitude at nine loops in planar N=4 super-Yang–Mills.
- The result release reports agreement across all 107,053 coefficients in a key comparison and zero violations in published samples of several constraints.
- Anthropic identifies the system as Claude Fable 5.1 running in Claude Science.
- Primary source: Anthropic's account of the nine-loop result.
A scattering amplitude is a mathematical object used to describe the probability structure of a particle interaction. Physicists often calculate it as a perturbative expansion: each “loop” is another layer of quantum correction. More loops can make an answer more precise, but they also make the space of possible expressions explode. The setting here is deliberately specialized. Planar N=4 super-Yang–Mills is a highly symmetric toy gauge theory, not a model intended to describe the world at a particle collider. Its value is that difficult methods can be tested in a setting with unusually rich mathematical structure.
The problem was not solved by summing a pile of diagrams one by one. Researchers use a bootstrap: begin with a large class of expressions that could be the answer, then rule almost all of them out with constraints from symmetry, limits, known singularities, and related quantities. An earlier eight-loop paper used “antipodal duality” to map an easier stress-tensor form factor into information about the six-point amplitude. Claude reportedly extended both a direct bootstrap route and the indirect form-factor route.
The analogy is reconstructing a complex machine from rules rather than finding its blueprint. You know the finished machine must fit through certain doorways, obey conservation rules, and match a few observable measurements. Each constraint removes thousands of wrong designs. The final calculation still requires meticulous symbolic software because one bad index, convention, or basis conversion can derail every later step. That software-engineering burden is precisely where the reported agent contribution lies.
Anthropic says researchers Liam Fitzpatrick and Siddharth Mishra-Sharma selected the challenge, supplied a short description, and periodically instructed the model to continue. The model built and debugged the implementation, used Python and SymPy in the direct route, organized classical computation, and delivered a result that could be checked. The company says a direct bootstrap used about 96 CPUs for a week and roughly $100 in classical compute; it estimates the end-user cost of either route at around $1,000 to $2,000, largely inference. Those figures are the authors' account, not a full published ledger of all prior model attempts, retries, or training cost.
The Cosmic9 release makes this story much stronger than a bare vendor claim. It reports rational-arithmetic certification at three primes for the form factor, separately calculated amplitude structure at two primes, matching direct and indirect representations, and agreement across 107,053 comparison coefficients. The prior eight-loop machinery also reproduced 1,000 randomly tested nonzero words without mismatch. The release reports checks on symmetry, vanishing conditions, entry restrictions, and related structure. These are not generic benchmark points; they are ways mathematical objects can contradict themselves.
Lance Dixon's role deserves precise framing. He did not publicly say he independently rewrote the whole project from scratch and blindly reproduced the final function. Rather, he mapped Claude's amplitude back to the related form factor his group had been studying for about two years and compared it with the expected structure and constraints. Anthropic quotes Dixon's positive assessment while he also emphasizes that the methods came from his collaborators' earlier work. That is meaningful expert cross-validation, but it is not the same thing as an unrelated group publishing a full independent replication.
Dixon's own forward-looking caution is the best quote to carry from the announcement: the major next test is whether AI systems “discover new physical principles before humans.” This result is not that. The agent worked inside a problem representation, a set of methods, and a validation culture built by people. The six- and seven-loop literature is part of the intellectual scaffolding that made nine loops a reachable target.
There is another reason not to oversell the milestone. A concurrent human-led effort by Song He, Jirong Jing, and Xiang Li released most of the nine-loop symbol around the same period through Zenodo. That does not diminish the agent's work, but it suggests the field had made the problem computationally ripe. The public programs are not distributed, and the complete function was only computed once; the release also notes assumptions in moving from symbol-level relations to function-level completion.
The strongest interpretation is therefore operational. Frontier models can now sustain a carefully scaffolded scientific workflow in which implementation details are scattered, error states are numerous, and success requires days of persistence. That can make an expert team faster even when the AI has not generated a new hypothesis. The correct follow-up is not “scientists have been replaced.” It is whether labs can reproduce this pattern on less symmetric problems, with open code, clearer failure accounting, and independent validation before the human literature already converges on the answer.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)