{"slug": "claude-beat-a-physics-world-record-for-2000-here-is-what-that-means", "title": "Claude Beat a Physics World Record for $2,000 — Here Is What That Means", "summary": "Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma used Claude Fable 5.1 to compute the nine-loop scattering amplitude in N=4 super Yang-Mills theory for roughly $1,000–$2,000 on 96 CPUs over one week, one loop past the eight-loop record Lance Dixon had held since 2023; Dixon verified the result and it matched on all 107,053 coefficients. The run used Anthropic's Claude Science harness with Python and SymPy across two independent methods, the bootstrap technique and an indirect form-factor approach, with the bootstrap step alone costing about $100. Song He's group at the Chinese Academy of Sciences independently reached the same nine-loop result at the same time using GPT-6 with a human-built framework.", "body_md": "Last week, two Anthropic physicists gave Claude Fable 5.1 a single task description, told it to report back every four to six hours, and walked away. One week later — running on 96 CPUs at a total cost of around $2,000 — Claude had computed the nine-loop scattering amplitude in [N=4 super Yang-Mills theory](https://www.anthropic.com/research/yes-claude-can-do-nine-loops). That is one loop past the world record Lance Dixon had held since 2023. Dixon verified the result himself, and it matched on all 107,053 coefficients.\n\n## What Was Computed and Why It Matters\n\nScattering amplitudes are the mathematical formulas physicists use to predict what happens when subatomic particles collide — the kind of calculations that back results from the Large Hadron Collider. The complexity grows steeply with each additional loop of calculation. Most amplitudes are computed to only two or three loops. Dixon’s eight-loop result in 2023 was a three-year project for a team of top physicists. Nine loops was considered too hard to attempt.\n\nLiam Fitzpatrick and Siddharth Mishra-Sharma, physicists at Anthropic, ran Claude through the computation via Claude Science — Anthropic’s paid research harness built on [Fable 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) that wraps the model in structured prompts for scientific use. The language stack was Python and SymPy. They ran two independent methods: the bootstrap technique (starting from all possible answers and eliminating until one remains) and an indirect form-factor approach. The bootstrap step alone cost roughly $100. The full run came to $1,000–$2,000.\n\n## The Autonomy Is the Actual News\n\nThe physics result matters. But the more significant signal is how little human involvement it required. Fitzpatrick and Mishra-Sharma essentially wrote a prompt and checked in every four to six hours. No intermediate scientific guidance. No hand-holding through the messy parts. Claude wrote all of its own Python code from scratch, debugged it autonomously, and produced a reliable result across both methods.\n\nDixon, who doubted the nine-loop result was even achievable, put it plainly: “The whole setup is very fragile: if you make any mistake at all in the computational recipe, it all crashes down like a failed soufflé.” Then he added: “Claude understands our 2019 and 2023 papers better than any human, aside from my co-authors.”\n\nCompare that to March 2026, when AI physics experiments looked more like graduate student work — small tasks, frequent corrections, limited scope. The physicist who [issued the original challenge](https://4gravitons.com), Matt von Hippel, framed the shift directly: “In March, AI was accomplishing physics projects like a student. In contrast, this is a real frontier calculation, the kind of thing normally tackled by the top experts.”\n\n## The Cost Collapse Is the Real Story\n\nA three-year team effort replaced by $2,000 and one week of autonomous compute is not just a physics story — it is a signal about where agentic AI is headed. Von Hippel’s key observation: “There is more low-hanging fruit out there than you’d expect.” Barriers that researchers attributed to mathematical difficulty were often just resource constraints — nobody had tried applying enough compute with enough organizational patience.\n\nAdding to the credibility: Song He’s group at the Chinese Academy of Sciences independently hit the same nine-loop result at the same time, using GPT-6 with a human-built framework. [Two independent teams, same result](https://aiweekly.co/alerts/anthropics-claude-computes-nine-loop-yang-mills-amplitude). No ambiguity.\n\n## What Developers Should Actually Do With This\n\nClaude Fable 5.1 is the model behind the computation — available now via the API as `claude-fable-5-1` at the same price as Fable 5, with 4x cheaper cache reads. Claude Science is the harness Anthropic used; it is a commercial offering, though details on developer access have not yet been published.\n\nThe more immediate takeaway is not about replicating particle physics. It is about the class of problem this unlocks: long-horizon, highly structured, computation-heavy tasks where the bottleneck has always been human patience and continuity. Compiling large codebases, auditing security at scale, running multi-day optimization passes — these are all problems where “keep going and report every four hours” is now a legitimate strategy.\n\nThe limitation worth acknowledging: Claude applied known methods with more compute, it did not derive new theoretical physics. Von Hippel was careful to note the uncertainty about whether this generalizes. But that is the floor of what this result demonstrates. The ceiling is still being mapped.", "url": "https://wpnews.pro/news/claude-beat-a-physics-world-record-for-2000-here-is-what-that-means", "canonical_source": "https://byteiota.com/claude-nine-loop-yang-mills-physics-2000-cost/", "published_at": "2026-09-27 04:11:54+00:00", "updated_at": "2026-09-27 04:29:33.996612+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-agents", "large-language-models"], "entities": ["Anthropic", "Claude Fable 5.1", "Claude Science", "Liam Fitzpatrick", "Siddharth Mishra-Sharma", "Lance Dixon", "Song He", "Chinese Academy of Sciences"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/claude-beat-a-physics-world-record-for-2000-here-is-what-that-means", "markdown": "https://wpnews.pro/news/claude-beat-a-physics-world-record-for-2000-here-is-what-that-means.md", "text": "https://wpnews.pro/news/claude-beat-a-physics-world-record-for-2000-here-is-what-that-means.txt", "jsonld": "https://wpnews.pro/news/claude-beat-a-physics-world-record-for-2000-here-is-what-that-means.jsonld"}}