What Claude's record physics calculation actually means

Anthropic says its Claude AI model has set a new record for one of the hardest kinds of calculations in theoretical physics — here's what that actually means, in plain terms, without the maths-department jargon getting in the way.

Physicists use something called a "scattering amplitude" to predict how particles behave when they collide with one another. These calculations come in stages called loops: the more loops involved, the more accurate the result but also the harder it is to compute. Claude worked out a calculation with nine loops, one more than the previous record of eight, set by physicists Lance Dixon and Andy Liu back in 2023.

The calculation used a simplified, made-up version of particle physics called N=4 super-Yang-Mills, which isn't meant to describe real particles at all. Physicists use it purely as a testing ground for developing new mathematical techniques they can later apply elsewhere, according to Anthropic.

It started life as a public challenge, posted on 7 August by physics writer Matt von Hippel to see what current AI models could manage. Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, gave Claude a single instruction and then mostly let it get on with the work entirely by itself.

Claude solved the problem two different ways, arriving at matching answers by two separate routes, taking about a week of computer time on 96 processors, at a cost Anthropic put at around $1,000 to $2,000 for each method — a genuinely small sum for a calculation of this difficulty.

A physicist at Stanford's SLAC laboratory, Lance Dixon, spent about two weeks checking the answer carefully to make sure it was right before anything was published. Meanwhile, a separate team in Beijing led by Song He arrived at nearly the same answer around the same time, using some help from OpenAI's GPT-6 model along the way for parts of its own working.

Anthropic isn't claiming this reveals anything new about the universe itself. Claude followed an already-known method for a very difficult calculation; the achievement is that it carried out that method correctly, largely unsupervised, not that it discovered new physics along the way.

If you're wondering why any of this matters beyond physics circles, the answer is really about trust: a result this technical, checked independently by a working physicist and largely matched by a separate team on the other side of the world, is harder to dismiss as a lucky guess or a cherry-picked demo than most AI claims you'll read about.

None of it changes what Claude can do for you day to day; it's a signal about how far AI can be trusted with narrow, well-specified expert tasks when someone who understands the subject sets it up carefully first.