Claude broke a physics record from a single prompt

This week @AnthropicAI posted a headline that reads like a dare answered: "Yes, Claude can do Nine Loops." It links to a guest post on Anthropic's science blog, and it's worth a few minutes even if you've never touched theoretical physics, because the interesting part is how the work got done.

What actually happened

Physicists predict how particles collide using formulas called scattering amplitudes. They build those answers in layers of finer corrections called loops, and each extra loop makes the answer more precise while costing a lot more computation. In a simplified model physicists use as a proving ground, planar N=4 super Yang-Mills, the record stood at eight loops, set by SLAC's Lance Dixon and collaborators in 2023.

On August 7, physicist turned science writer Matt von Hippel posted a public challenge to AI companies on his blog, 4gravitons: push one of these calculations a loop further, using only the kind of computing budget an academic could get.

Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma took it on. According to Anthropic's write-up:

• The starting prompt was one sentence naming the problem: the six-particle amplitude at nine loops.
• The follow-up instructions amounted to "keep working until I tell you to stop, and send updates every few hours."
• Claude Fable 5.1 ran largely unsupervised for days inside Claude Science, Anthropic's paid platform for researchers, and solved it two different ways.
• Total cost landed around $1,000 to $2,000. One of the two methods used about $100 of compute, roughly 96 CPUs running for a week.
• Dixon spent about two weeks checking the result and confirmed it.

Anthropic also points to a parallel result. A group at the Chinese Academy of Sciences led by Song He published most of the same nine-loop result this month with GPT-6 as an assistant, though humans built the overall framework there.

The honest read

Von Hippel is careful about what this means, and the rest of us should be too. Claude did not invent new physics. It ran methods Dixon and his colleagues had already developed, with more compute and more patience than a grad student could give it. Von Hippel's own summary is that this was "known methods with somewhat more compute." He also says the tools have genuinely improved, since Claude got there with nothing fancier than repeated instructions to keep going.

I think both halves matter. The record is real, and the recipe and the final check still came from people.

The check is the part headlines skip. The calculation took days, and verifying it took an expert two weeks. Producing the answer wasn't the slow step. Knowing whether it was right was, and very few people alive were in a position to judge.

Why an owner in the Fox Valley should care

You are not computing gluon amplitudes. Still, the shape of this result matches how AI is starting to do real work inside ordinary businesses.

Claude succeeded because a precise, documented method already existed, and it followed that method further than people had time to. In a business, the equivalent is a written process: how you qualify a lead, how you build a quote, what goes into the month-end report. AI can run a clear recipe for as long as you let it. It can't reliably run one that lives only in your office manager's head.

"Keep going until I say stop, report back every few hours" is also a management style, not a chat. Agents that grind on a task for hours or days are showing up in everyday tools, so you have to decide what they report and who reads it.

Then there's the checker. Dixon's two weeks of verification is the piece worth copying. Any automation that produces work on its own needs someone who knows what right looks like, plus a way to spot-check without redoing everything by hand.

Cost, meanwhile, has stopped being the excuse. A few thousand dollars bought a physics record. Having AI work through a well-defined task keeps getting cheaper, and the expensive part is defining the task well.

Where to start

If you want AI doing multi-step work in your business, start where Anthropic's physicists did: a recipe that's already written down and a person who can tell when the output is wrong. Most small businesses have the person. Far fewer have the recipe on paper.

That gap is what our free process audit looks for. We walk through where your team's time goes, find the processes clear enough to hand to automation, and flag the ones that need writing down first.