Anthropic Says Claude Pushed a Physics Calculation Past the 2023 Human Record, One Loop Higher

Since physicist and blogger Matt von Hippel dared AI companies on August 7 to crack one of two unsolved amplitude problems using an academic-sized compute budget, Anthropic says its Claude model cleared the harder bar within weeks.

Anthropic published the result on its Science Blog on September 25, in a post written by von Hippel himself. According to Anthropic and Unite.ai, two of the company's physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, contacted von Hippel at the end of August to say they had tackled his challenge and wanted him to check the work with SLAC's Lance Dixon before going public.

What Claude actually calculated

The target was the six-particle scattering amplitude in planar N=4 super Yang-Mills theory at nine loops. Scattering amplitudes are formulas physicists use to predict what happens when particles collide. Loop order measures how many rounds of quantum corrections get folded into the answer, and each added loop multiplies the algebra involved, often exponentially, according to Alpha Signal and Unite.ai.

N=4 super Yang-Mills is not a description of the real universe. It's a deliberately simplified "toy" theory researchers use to test methods that are too hard to develop directly against real particle physics, per Unite.ai's account of von Hippel's writing. The previous record in this specific six-particle calculation was eight loops, set by Dixon and Yu-Ting Liu in 2023, according to kingy.ai (Anthropic's own account and other outlets render the collaborator's name as Andy Liu).

Anthropic says Claude, running inside a specialized "Claude Science" harness on a system referred to as Fable 5.1, was given a single prompt describing the nine-loop problem and then worked largely without human correction for several days. HuggingNews and Alpha Signal put the total compute cost at roughly $1,000 to $2,000. Anthropic's own post on X describes it only as "a total cost of a few thousand dollars."

Claude reportedly solved the problem two separate ways: a direct bootstrap of hexagon functions, and an indirect route through the nine-loop form factor using a technique called antipodal duality. Both paths agreed, which is the standard method researchers use to catch errors. Dixon then independently verified the result using his own group's tools, according to Anthropic, kingy.ai and Crypto Briefing.

The caveats worth taking seriously

Anthropic's announcement is a guest post, and Unite.ai's account is explicit that the arrangement has a wrinkle: Anthropic invited von Hippel to write it, compensated him for his time, and staff gave feedback on drafts, though the company says the content and opinions are his own. That doesn't make the physics wrong, but the primary writeup of the result comes from someone Anthropic paid, not an independent journalist or a peer-reviewed paper.

Runtime Wire raised the sharpest technical pushback. It notes that a 2016 paper already reported a nine-loop result for a related five-point amplitude in the same theory, meaning the loop count by itself isn't unprecedented. What's new here specifically is the six-particle version. Runtime Wire also argues that showing a model can reproduce a known type of expression is a different capability than deriving a genuinely new one, and both differ from a result that gets fully, independently reproduced by outside researchers.

Kingy.ai adds a more technical wrinkle from inside Anthropic's own released files: there's a distinction between the "symbol," a mathematical representation that's been extensively cross-checked, and the fuller function, whose independent verification is described as less complete. Kingy.ai says it inspected the linked documentation and ran limited file and arithmetic checks but did not reproduce the calculation or independently validate the underlying physics itself.

Alpha Signal reported one detail not corroborated elsewhere in the source set: that a separate human research team using OpenAI's GPT-6 as an assistant reached a similar result within days of Claude's run. No other outlet in this review mentions that parallel effort, so it should be treated as a single, unconfirmed claim rather than an independently verified race between labs.

Where this leaves the record

Nobody involved is calling this a new law of physics. Dixon, Fitzpatrick, Mishra-Sharma and von Hippel all frame it as a demonstration that known bootstrap techniques, given enough sustained computation, can be pushed one loop further than a human-led team managed in 2023. The open question is whether outside physicists, working from Anthropic's released files independently of the company's own writeup, will confirm the fuller function to the same standard the symbol has already met.