Anthropic says Claude solved a nine-loop physics problem for about $2,000. A Chinese team had already posted the same answer.

Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, gave Claude a one-line problem statement in late August: compute the six-particle ("hexagon") scattering amplitude in planar N=4 super Yang-Mills theory at nine loops -- one loop past the previous record. A week and about $2,000 of compute later, Claude had produced the answer two different ways. Lance Dixon, the SLAC National Accelerator Laboratory and Stanford physicist who set the prior eight-loop record in 2023, checked the math himself and confirmed it holds.

N=4 super Yang-Mills is not a theory of anything that exists. Physicists use it as a sandbox: it shares deep mathematical structure with the equations that describe real particle collisions, but it's clean enough to actually solve. A "loop" is a unit of calculational difficulty -- each one adds a layer of quantum correction, and the arithmetic needed to track them grows, in von Hippel's words, "exponentially or even factorially." Nine loops was, until this month, past the edge of what anyone had directly computed.

The occasion was a dare, not a research agenda. On Aug. 7, physicist-turned-writer Matt von Hippel posted a public challenge to AI companies: solve one of a short list of the field's genuinely hard open problems, using only "the kinds of computer resources an academic has access to." His stated bet was skeptical -- that AI would keep failing this test the way it had failed others, until, predictably, someone's field was the one it finally cracked. Nine loops of N=4 super Yang-Mills was one of two problems on his list.

Anthropic's setup, which it calls Claude Science, ran Claude driving Python and the symbolic-math library SymPy across the equivalent of 96 CPUs for about a week, with the two physicists checking in roughly every four to six hours rather than steering each step. Claude worked the problem two ways: the original bootstrap method physicists have used for this class of calculation since the 2010s, and a second, independent route through "form factors" that cross-checked the first.

This isn't Anthropic's first attempt at pointing Claude at open physics problems with a light hand on the wheel -- the company ran a related "vibe physics" effort earlier this year testing how far a model could get on research-grade problems with minimal scaffolding. The nine-loop run is the same bet at a harder, more externally-verifiable target: a problem with a named expert, a specific numeric answer, and someone with the standing to say whether it's actually right. Anthropic breaks the cost down further: about $100 of that went to the core bootstrap calculation alone, with the rest of the roughly $1,000 to $2,000 total spent on the independent cross-check and the exploration around it.

“The whole setup is very fragile: if you make any mistake at all in the computational recipe, it all crashes down like a failed soufflé.” — Lance Dixon, SLAC / Stanford, on verifying Claude's result

Dixon's own reaction undercuts any read of this as routine. He told Anthropic the direct bootstrap approach to nine loops was one he'd considered "too hard to do directly" -- his own 2023 eight-loop record used an indirect route through form factors and a duality relation instead. (Dixon didn't just bless the number -- he said Claude "understands our papers better than anyone else," which is a different and larger claim than getting one calculation right.)

That last date matters, because Anthropic's post does not stand alone. Song He, a scattering-amplitudes researcher at the Chinese Academy of Sciences in Beijing, says his group already had the majority of the same nine-loop result -- using AI assistance from GPT-6 to help fix constraints within a process two humans were actively directing, not the largely self-steered run Anthropic describes. He, Jirong Jing and Xiang Li posted their dataset, "The Symbols of Six-Gluon MHV Amplitudes through Nine Loops," to Zenodo on Sept. 17 -- eight days before Anthropic's post went up.

That distinction is the actual news, not a footnote to it. A model producing a correct, expert-verified physics result with minimal supervision is a different capability claim than a model helping human experts go faster at something they were already doing -- even if the two processes land on the same number in the same month. Conflating the two is exactly the kind of framing gap that makes a genuine capability jump hard to tell apart from a well-timed press cycle.

For a reader tracking what Claude Fable 5.1 and its peers can actually do unsupervised, the honest summary is narrower than "AI solves physics problem": a frontier model, checked in on twice a day, executed a known-but-brutal calculation correctly on the first fully-reported attempt, on a task its own verifier didn't expect to be tractable that way. Whether that generalizes past one physicist's dare is the open question the apply items below are actually about.

It's also a small, telling data point for Anthropic's own positioning in the frontier-lab race: a capability demo that costs a few thousand dollars and produces a result an outside expert is willing to stake his own name on is a cheaper, harder-to-fake credibility signal than a benchmark score the company graded itself. Whether rivals answer with their own version of von Hippel's dare -- rather than another self-reported eval -- is worth watching the same way the reasoning model race itself is.