Anthropic published a guest post by physicist-turned-science-writer Matt von Hippel, reporting that two of its researchers used Fable 5.1 inside Claude Science to compute a nine-loop scattering amplitude in a toy particle-physics theory (planar N=4 super Yang-Mills) — a calculation von Hippel had publicly challenged AI companies to attempt a month earlier on his blog, 4gravitons.com. Anthropic's physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, gave the model a one-line prompt and periodic "keep going" instructions; the result was checked by Lance Dixon (SLAC/Stanford), who had computed the previous, eight-loop version of the same formula in 2023.
What the primary source itself actually concludes
This is worth stating plainly, because it cuts against how the story reads at a glance: von Hippel's own summary is that nothing surprising happened, computationally. "It did something it turned out humans were also able to do. Claude used known methods, with a bit more compute than people had tried to use before... not super-intelligently so." He'd hoped to see an AI overcome a limit computationally, in a way that would tell him something about future risk; instead, he concludes the field's outstanding problems were more tractable with ordinary methods and money than researchers assumed, not that a new kind of intelligence broke through a wall. Lance Dixon's own addendum agrees: "it used all the methods my collaborators and I developed over the years," calling it "a triumph... to execute all of the steps in the complicated recipe we laid out," not a display of new physical insight — with that distinction reserved for a future he says hasn't arrived yet.
Secondary coverage has run further than that: outlets picked this up under headlines like "Claude Breaks Physics Record" and "Claude Solves 9-Loop Physics Problem to Break SLAC Record." Neither the guest post nor Dixon's addendum uses "record" or "breakthrough" — the loop count is a new highest for this specific formula, but the authors' own framing is "known techniques, more budget," which is a meaningfully smaller claim than what's circulating around it.
A Chinese team published a closely related result first — not "a few days after"
Von Hippel's post says Song He's group at the Chinese Academy of Sciences got the majority of a related quantity "a few days after I heard from Anthropic," using GPT-6 for some constraints. Checking the actual record: Song He, Jirong Jing, and Xiang Li published a dataset on Zenodo, "The Symbols of Six-Gluon MHV Amplitudes through Nine Loops," on September 17, 2026 — eight days before Anthropic's post went live on September 25. Anthropic's team reached Dixon for verification around September 1, so both efforts were underway in roughly the same window, and von Hippel's story is honest about the concurrency rather than claiming Claude got there alone. But "published first" belongs to the Chinese Academy of Sciences team, publicly and on the open record, not to Anthropic — a distinction that framing built around "then people did it" a few paragraphs after Anthropic's own effort tends to obscure.
The disclosure is worth reading, not skipping
The post ends with a straightforward disclosure: Anthropic invited and paid von Hippel to write it, gave feedback on drafts ("the content and opinions are his own"), and gave Lance Dixon Claude usage credits for doing the independent verification. That's a materially different arrangement from an unaffiliated journalist or an uncompensated academic peer review — it doesn't make the physics wrong, and von Hippel's piece is candid enough to undercut Anthropic's own framing in several places, which is itself evidence the arrangement didn't buy uncritical coverage. But it's a compensated guest post and a compensated verification, not independent third-party confirmation, and should be read as such.
Where this fits against the bigger claim
This lands three days after Dario Amodei's tweet argued biology should be expected to follow the same trajectory as AI-driven math progress, pointing to "beginning to solve the top few open problems in all of mathematics" as evidence of a trend. This result is a genuinely relevant data point for that trend — real execution of a hard, well-defined computation at a new scale, for a few hundred to a couple thousand dollars. But it's also a useful illustration of the distinction that matters: this is known technique plus more compute solving a well-posed problem with a checkable answer, which is exactly the category von Hippel's own conclusion says it falls into — not new mathematical ideas, and not the kind of open, contested claim this blog flagged in the disputed OpenAI Navier-Stokes case. Progress of this specific, verifiable kind is real; it's a narrower claim than the sweeping one it's being cited to support.