Yes, Claude Can Do Nine Loops: Inside Anthropic's Amplitude Result

Update — September 26, 2026: This post has been rewritten against Anthropic's full published article, "Yes, Claude Can Do Nine Loops", to correct earlier speculative details (model identity, cost breakdown, and the Song He comparison) with what Anthropic and physicist Matt von Hippel actually reported.

Theoretical physicists predict how particles behave using formulas called scattering amplitudes. They are notoriously hard to compute, so researchers build the answer up in layers of increasingly fine corrections called loops — each added loop sharpens the prediction but multiplies the computation. Most real calculations stop at two or three loops. Eight loops was the previous record in planar N=4 super Yang-Mills, a simplified "toy model" physicists use as a testing ground, set by Lance Dixon of SLAC National Accelerator Laboratory and collaborators.

On August 7, 2026, physicist-turned-science-writer Matt von Hippel (better known online as @4gravitons) issued a public challenge: could an AI lab push past eight loops in this model — using only the kind of compute budget an academic could reasonably access? Given a single prompt describing the nine-loop problem, Claude ran largely unsupervised for days inside Claude Science and solved it, using the same bootstrap methods Dixon and his colleagues had developed, at a total cost of a few thousand dollars. Dixon independently verified the result, and von Hippel wrote up the experience for Anthropic's blog.

If you build agents, the interesting part is not the loop count alone. It is the operating pattern: a multi-day, largely unsupervised run, a problem where "wrong" is detectable through known physical limits, and a domain expert who could actually catch a mistake — the same ingredients that made Claude's Riemann zeta bound extension credible without touching the full hypothesis.

TL;DR — what practitioners ask first

What the nine-loop hexagon amplitude actually means

Scattering amplitudes answer a concrete question: if you collide particles with specified properties, what is the quantum-corrected probability that a given outcome happens?

For six particles in the MHV (maximally helicity violating) configuration, the tree-level answer has been known since Parke and Taylor's 1986 formula — a compact rational function. Each loop correction adds integral functions of increasing complexity, and the number of integral topologies explodes with loop order. By nine loops, physicists lean on planarity, integrability, and years of community-built reduction identities (symbol alphabets, the bootstrap method) just to keep the ansatz tractable at all.

N=4 super Yang-Mills is not the theory that describes protons or the particles in the Standard Model. It is a maximally supersymmetric "sibling" theory that strips away complications real QCD carries, leaving a mathematically clean laboratory where amplitudes often collapse into structured polylogarithmic functions. Physicists use it to develop and stress-test techniques — like the bootstrap method Dixon's group pioneered — that later inform harder, more realistic calculations.

So when Anthropic says "nine loops," the achievement is that Claude executed Dixon's own established recipe correctly and consistently for long enough to produce a final expression that survives physical consistency checks — not that it discovered a new physical principle.

What actually happened, according to Anthropic

Von Hippel's original post asked for either "N=8 supergravity to seven loops, or N=4 super Yang-Mills to nine loops," specifically framed around using the kind of computer resources an academic has access to — a direct dare to AI labs that raw compute spend was not the point.

Anthropic researchers Liam Fitzpatrick and Siddharth Mishra-Sharma took up the challenge inside Claude Science, Anthropic's workbench for scientific research (explainx.ai covered the platform's launch here). They gave Fable 5.1 a single prompt describing the nine-loop hexagon amplitude problem, then — in von Hippel's own words — told it something close to: "I'm going to sleep and won't be available for another several hours. Keep working on this until I tell you to stop. Give me updates every 4-6 hours."

Claude worked the problem two independent ways:

1. The bootstrap method — the direct, original approach Dixon's group has used for years: propose a candidate function space, then constrain it against known physical limits (collinear and soft factorization, lower-loop data) until only one answer survives — a process von Hippel compares to solving a very hard Sudoku.
2. An indirect form-factor approach — reaching the same object via antipodal duality, a separate mathematical relationship that gives an independent cross-check on the first method.

Both were implemented in Python with SymPy. Von Hippel notes Claude completed the calculation "in one shot, without any scientific oversight more sophisticated than keep going" — a level of unsupervised persistence that would typically take a human researcher several false starts to match.

Why Lance Dixon's sign-off carries weight

Lance Dixon is one of the central figures in modern amplitude technology, from early multi-loop QCD work through the development of the symbol methods and bootstrap constraints that planar N=4 calculations depend on today. He held the prior eight-loop record in this exact model — which made him the one physicist positioned to know immediately whether Claude's nine-loop answer was right.

Dixon independently verified the result and confirmed it holds up in the same formats used for previous loop orders. In his own addendum to Anthropic's post, he described the calculation as "fragile" — a single error anywhere in the chain would collapse the entire result — and credited Claude with correctly executing "the complicated recipe" end to end. He was also careful to frame the ceiling: the genuinely transformative moment, in his view, will come "when large language models start to come up with new physical principles," not when they execute existing ones very well.

The compute bill — and what "a few thousand dollars" means

Anthropic put the budget for each of the two methods at roughly $1,000–$2,000. The bootstrap calculation — the one that actually produced the verified nine-loop result — used only about $100 of that, which Anthropic describes as computationally equivalent to 96 CPUs running for one week. That number matters because it was von Hippel's actual condition for the challenge: not "can AI solve this with unlimited compute," but "can AI solve this with what an academic group could reasonably access."

It excludes, notably:

• Dixon's own validation time.
• The engineering behind Claude Science's harness.
• Anthropic's cost of the parallel form-factor approach, which did not end up being the one that landed the headline result.

Compare that to the historical baseline: a nine-loop amplitude push has typically meant person-months of dedicated expert effort. Even generously loaded, a few-thousand-dollar compute bill reframes what "expensive" means for this class of problem — a pattern explainx.ai also saw in multi-agent math swarms tackling Navier-Stokes, where independent reproducibility mattered more than the sticker price.

Song He's concurrent result — confirmation, not a contest

Within roughly two weeks of Claude's run, Song He's group at the Chinese Academy of Sciences independently reached a related nine-loop form-factor result, reportedly with help from GPT-6-class tooling. Anthropic and von Hippel treat this as mutual confirmation rather than a race: two separate teams, using two different AI systems and slightly different mathematical routes, converged on frontier results in the same model within weeks of each other.

That framing matters more than a leaderboard would. The takeaway is not "Claude beat GPT-6 to nine loops" — it's that multiple independent groups now have the tooling to push past a record that had stood since Dixon's original eight-loop work, which is a stronger signal than either result alone.

Builder angle — what to copy from this workflow

If you are not a physicist, the transferable pattern looks like this:

1. Pick problems with cheap, physical falsifiers

Nine-loop amplitudes fail loudly when collinear and soft limits break. Riemann bounds fail when a Lean check fails. Agent loops without a falsifier drift; agent loops with one can run for days on a single prompt.

2. Give the model a long leash and a check-in cadence, not a script

Von Hippel's actual instruction was closer to "keep going, update me every few hours" than a detailed procedural plan. The value came from Claude sustaining Dixon's known method correctly over a long run — not from novel instructions on every step. See loop engineering for coding agents for the same pattern applied to software.

3. Run two independent methods when you can afford it

Claude's bootstrap and form-factor approaches served as cross-checks on each other before Dixon ever saw the result. That is the science-run equivalent of two independent CI pipelines — cheap insurance against a single silent error in a fragile calculation.

4. A domain expert still has to sign off

$100–$2,000 in compute does not remove the need for someone like Dixon, who could have caught the mistake, to actually check the work. Budget reviewer time even when tokens are cheap, especially before you publish.

5. Don't conflate "hard calculation" with "new physics"

Dixon's own framing is the right one to borrow: executing an established recipe extremely well is not the same as originating a new physical principle. This result belongs next to incremental amplitude milestones, not next to Millennium Prize headlines — a distinction explainx.ai also draws in its Millennium Prize fact-check.

What people are asking

Is nine loops a world record? Yes, in the specific sense that matters here: it is one loop past the previous published record (Dixon's eight loops) for the planar N=4 SYM six-particle MHV amplitude, independently verified by the physicist who held that prior record.

Could I reproduce this with public Claude? Fable 5.1 running inside Claude Science is the same family of model available more broadly, but this specific run relied on von Hippel and Dixon's domain expertise to frame the problem correctly and check the output — the model alone did not "discover" the target or validate its own answer.

Does this help my coding agent project? Indirectly, yes. The pattern — a well-specified problem with physical falsifiers, a long unsupervised run with periodic check-ins, and independent cross-checks — mirrors solid production agent design, even outside physics.

What's the actual takeaway, in von Hippel's own words? That there was "more low-hanging fruit out there than you'd expect" — the field had assumed compute and complexity were the bottleneck, when in practice sustained, disciplined execution of already-published methods got there first.

Honest limitations

• Not a new physics discovery — Dixon is explicit that this is excellent execution of an existing recipe, not a new physical principle.
• Planar N=4 SYM ≠ real-world QCD — phenomenology for actual particle collisions still needs separate techniques.
• Mass gap untouched — do not merge this story with Millennium Prize Yang-Mills headlines; see the fact-check.
• Cost figures are Anthropic's own estimates — a rerun on public pricing, with retries, could differ.
• Song He's result is reported via contemporaneous coverage, not a joint Anthropic publication — read both primary write-ups before treating the comparison as settled.

Related reading on explainx.ai

• Millennium Prize Problems and AI — fact-check hub (amplitude progress vs mass-gap proofs)
• Claude pushed a Riemann zeta bound from 41.6% to 67.2% with 60 subagents
• OpenAI's Navier-Stokes swarm — credit, disputes, and verification lessons
• Claude Science: Anthropic's AI workbench for scientists
• Claude Team plan for scientists — 10,000 free and discounted seats
• Loop engineering for coding agents (Claude Code patterns)
• Claude Code loops — /goal, schedules, and turn boundaries
• Fields Medalists' declaration on AI labs and math alignment

Official source: Anthropic — "Yes, Claude Can Do Nine Loops"

Facts, quotes, and figures in this post reflect Anthropic's published article as of September 26, 2026. Loop-order records and community acceptance may evolve as other groups review or extend the result.