On September 25, 2026, Anthropic published a guest post by physicist-turned-science-writer Matt von Hippel describing how Claude met a challenge he had posted on his blog about a month earlier: show that an AI, on the kind of computing budget an academic has, can solve one of the scattering-amplitudes field’s big outstanding problems, such as N=8 supergravity to seven loops or N=4 super Yang-Mills to nine loops. Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma took the second. Using Claude Fable 5.1 inside the Claude Science harness, they asked for the six-particle (hexagon) amplitude in planar N=4 super Yang-Mills at nine loops and, beyond that, mostly told it to keep going.
Claude did the calculation two ways, with the original bootstrap method and with the indirect form-factor approach that SLAC’s Lance Dixon had used to reach eight loops. Either route would have cost an end user around one to two thousand dollars, mostly in model usage; the bootstrap computation itself, written in Python with SymPy, used about 100 dollars of that, equivalent to running 96 CPUs for a week. Dixon, who checked the result, writes in an addendum that he was told on September 1 that Claude had computed the nine-loop MHV six-particle amplitude and was asked to validate it.
Von Hippel is careful about what this proves. Claude used known methods with somewhat more compute than people had tried before, not a new technique; Dixon notes it used the methods his collaborators developed. And a few days after Anthropic’s result, Song He’s group at the Chinese Academy of Sciences reported it had already obtained most of the same result, with GPT-6-based assistance used less autonomously. The humans will publish and analyze the amplitude; N=4 super Yang-Mills is a toy theory used to hone techniques, not a description of the real world.
Why it matters: a frontier calculation normally tackled by the field’s top specialists was completed end to end by a harnessed model with oversight no more sophisticated than “keep going,” on a modest budget, and verified by the expert best placed to check it. What it does not show is AI breaking a computational barrier by new insight. Von Hippel’s own takeaway is that there was more low-hanging fruit than experts assumed, and that a human group was close behind.