Claude has computed a scattering amplitude that theoretical physicists had never reached: the six-particle amplitude in planar N=4 super-Yang-Mills theory at nine loops. According to Anthropic, two of its physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, used Claude to get the result with very little human input. The whole project cost somewhere between $1,000 and $2,000.
It’s a narrow result in a very specialized corner of physics. But it’s one of the clearest cases so far of an AI model finishing a frontier research calculation that experts had publicly named as an open challenge.
🎯 The challenge behind it
On August 7, 2026, former physicist and science blogger Matt von Hippel issued an open challenge to AI companies. He wanted proof that an AI working with academic-level computing resources could solve one of the big open problems in the scattering amplitudes field. He named two targets:
- Work out whether N=8 supergravity diverges at seven loops
- Find the six-particle amplitude in N=4 super-Yang-Mills at nine loops
Claude took on the second one. Anthropic published the result on September 25, less than two months after the challenge went up.
🔬 What “loops” actually means
Physicists use formulas called scattering amplitudes to predict what happens when particles collide. They can’t get these answers in one shot, so they build them from layers of smaller and smaller corrections called “loops.” Each extra loop makes the answer more precise.
It also makes the math much harder. Von Hippel pointed out that these calculations tend to grow exponentially, or even factorially, with each loop you add. His point was that researchers wouldn’t need new ideas to solve these problems, just far more computing power and effort than a typical academic group has.
⚙️ How Claude did it
Anthropic says the team used Claude Science, a paid platform that wraps the Claude model in structured rules and prompts so it behaves more reliably on scientific work. Claude reached the answer in two separate ways:
- Direct bootstrap. It built the amplitude from “hexagon functions,” a known mathematical toolkit for this problem.
- Indirect route. It worked through the nine-loop form factor, then used a technique called antipodal duality to convert that into the amplitude.
The two answers matched. In this field, getting the same result from two independent paths is the standard way to check for mistakes.
💰 The numbers
- Total cost: roughly $1,000 to $2,000
- Bootstrap compute: about $100, which works out to 96 CPUs running for about a week
- Code: Python with the SymPy symbolic math library
- Time from challenge to announcement: about seven weeks
The cost stands out most to me. You don’t need a supercomputer allocation or a big grant for this. It’s closer to a conference travel budget.
✅ Independent verification
On September 1, the Anthropic team sent the result to Lance Dixon, a particle physicist at SLAC and Stanford, and asked him to check it. Dixon spent the next two weeks validating it, mostly through the nine-loop form factor, a result his own team had been working toward for a couple of years.
That matters. A claim like this means much more once a leading expert in the field has independently confirmed it.
⚠️ The limitations
It’s worth being clear about what this isn’t. Claude didn’t discover new physics or invent a new method. It used bootstrap techniques that human physicists had already developed. Von Hippel set up the challenge that way on purpose, to test whether AI can carry out a known but exhausting calculation.
So this shows AI can grind through hard, well-defined problems, not that it has scientific creativity. It’s also one domain, one problem, and one result. We don’t yet know how well this approach carries over to messier research questions.
💡 Why it matters for practitioners
If you run technical or research-heavy work, here’s what to take from it:
- Look at your backlog. Many teams have “we know how, we just don’t have time” problems. Those are the kind this approach handles well.
- Build in two independent checks. Claude’s two methods agreeing did more for credibility than any single run could.
- Get outside review. Anthropic still needed Dixon to confirm it. AI output at this level still needs an expert to check it.
- Recheck your cost assumptions. Work that once seemed to need major funding may now fit in a small budget.
The next question is whether AI can go beyond executing known methods and start proposing new ones. Von Hippel’s other challenge, the seven-loop supergravity question, is still open. Anthropic’s full write-up has more technical detail on both calculation paths.