Anthropic turns a physics record into a marketing case for Claude Science

Anthropic used a physics record to make its argument for Claude as a research tool, saying the model computed a nine-loop six-particle scattering amplitude beyond the previous eight-loop record — a technical feat that also doubles, quite usefully, as marketing for its Claude Science offering aimed at scientific and technical customers.

The calculation, in a simplified model called planar N=4 super-Yang-Mills, has no direct commercial application of its own, but it gives Anthropic a hard, independently checkable claim to point to as it competes for scientific and technical users against rivals including OpenAI, according to Anthropic's own account of the work.

Claude answered a public challenge posted on 7 August by physics writer Matt von Hippel; Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma gave the model, running as Fable 5.1 inside Claude Science, one instruction and largely let it run unsupervised from that point onward.

Anthropic said the model completed the calculation two separate ways, each taking about a week on 96 CPUs at a cost of roughly $1,000 to $2,000, a modest bill for a record-breaking calculation.

SLAC physicist Lance Dixon spent about two weeks independently verifying the result before publication. Notably, a separate team led by Song He in Beijing reached most of the same answer around the same time using GPT-6-based AI assistance for some of its own constraints, meaning a rival AI system was already being pointed at the same class of problem too.

Dixon's own assessment, that the model correctly executed every step of the complicated recipe involved, is the kind of independent endorsement Anthropic did not have to manufacture itself, and one that competitors will find considerably harder to dismiss than an internal benchmark score alone.

Anthropic was careful to say the achievement is not new physics — Claude executed an established calculation method rather than discovering anything about how particles actually behave — a distinction that matters greatly for how much weight rivals and academics should ultimately place on the claim.

Even so, the timing helps Anthropic's pitch: as OpenAI and Google both push their own models toward scientific and technical use cases, an independently verified physics result is a more concrete data point than most companies in this race currently have to offer prospective customers.

The low cost is itself part of the pitch: at $1,000 to $2,000 per method and about a week of computing on 96 CPUs, the run shows how cheaply a well-specified expert task can now be handed to a model, a point analysts assessing AI research tools are likely to weigh.

How much of that framing survives scrutiny will depend partly on whether other labs can reproduce comparably hard results on demand, rather than on a single well-publicised case chosen, in part, because a public challenge already existed for it.