Claude Computed a Nine-Loop N=4 SYM Amplitude — SLAC Checked It

Planar N=4 super Yang-Mills is the cleanest strong-coupling laboratory quantum field theorists have: amplitudes often collapse to polylogarithms, symmetries are sharp, and loop orders that would be hopeless in QCD become a structured ladder of integral reductions. Getting the six-gluon MHV amplitude right at nine loops is not a viral "AI solved physics" headline — it is a precision stress test for whether an agent stack can sustain thousands of consistent symbolic steps without drifting conventions.

In September 2026, Anthropic published a research note saying an orchestrated Claude run derived that nine-loop object, exported it in a form Lance Dixon's group at SLAC could feed into their established validation pipeline, and received a matching sign-off after independent reproduction. Anthropic quotes total model-side compute on the order of $1–2k. The same month, Song He's collaboration described overlapping amplitude work assisted by GPT-6-class models on a nearby loop target with more explicit human algebra steering.

If you build agents, the interesting part is not the loop count alone. It is the operating pattern: a multi-day run, explicit checkpoints, a second-party verifier with different code, and a problem where "wrong" is detectable through factorization limits — the same ingredients that made Claude's Riemann zeta bound extension credible without touching the full hypothesis.

TL;DR — what practitioners ask first

What nine-loop six-gluon MHV actually means

Scattering amplitudes answer a concrete question: if you collide particles with specified helicities, what is the quantum-corrected probability amplitude?

For six gluons in the MHV configuration, the tree-level answer has been known since Parke and Taylor's 1986 formula — a compact rational function. Loop corrections add integral functions of Mandelstam invariants. Each loop order increases the integral topology count explosively; by nine loops, humans typically rely on planarity, integrability hints, and years of community-developed reduction identities to shrink the ansatz.

N=4 super Yang-Mills is not the theory inside protons. It is a sibling model where supersymmetry removes many of the complications that make real QCD intractable. Physicists use it because the same mathematical machinery (polylogarithms, symbol alphabets, bootstrap constraints from physical limits) trains techniques that later migrate toward phenomenology.

So when Anthropic says "nine loops," read it as: the agent maintained a convention-stable reduction chain long enough to emit a final function that survives collinear, soft, and lower-loop regression checks — not that Claude invented a new physical law.

What Anthropic says Claude did in September 2026

Anthropic's note (September 2026) describes a Claude Code–style orchestration: the model proposed integral topologies, called computer algebra tools, rewrote integrands when factorization tests failed, and checkpointed intermediate bases when token context threatened to truncate reasoning.

Three details matter for builders:

1. Convention lock-in early. The run fixed a trace basis, variable naming, and planar limit definition before loop seven — late convention drift is how amplitude automation silently diverges.
2. Physical limit unit tests. After each major reduction stage, the stack evaluated collinear and soft limits against known split functions. Failures triggered rollback to the last good checkpoint rather than "keep going."
3. Export for humans. The final artifact was not a PDF claim — it was a machine-readable expression plus a reproducibility bundle SLAC could ingest without Anthropic's internal harness.

That third point is why the SLAC validation lands differently from a standalone model transcript. Dixon's group did not re-trust the agent; they re-ran their own reduction stack and compared.

SLAC validation — why Lance Dixon's sign-off carries weight

Lance Dixon is among the central figures in modern amplitude technology — from early multi-loop QCD work to the integration of symbol methods and bootstraps that modern planar N=4 calculations depend on. His SLAC group maintains reduction pipelines that the wider community uses as informal ground truth.

Anthropic reports that after export, SLAC:

• Re-derived the nine-loop expression with independent code paths.
• Matched Anthropic's function up to known basis choices.
• Verified factorization constraints and regression against published eight-loop data.

This is not Clay Institute certification. It is the kind of external reproduction physicists treat as meaningful progress — similar in spirit to outside experts examining Claude's Riemann bound, but with a community that already shares machine-readable amplitude formats.

The $1–2k compute number — and what it omits

Anthropic's $1–2k figure is model/API spend for the orchestrated run, not fully loaded science cost. It excludes:

• SLAC validator time (human expert hours).
• Engineering to wire tools and checkpoints.
• Opportunity cost if the run fails at loop eight and must restart.

Still, compare orders of magnitude: a nine-loop amplitude push historically meant person-months of focused expert work. If the cited band holds for reproducible external validation, it reframes agent budgets from "chat spend" to "cheap iteration on hard symbolic search."

That framing aligns with how explainx.ai covers multi-agent math swarms — headline compute hours matter less than whether independent reproduction is built into the workflow.

Song He, GPT-6 assistance, and the contrast Anthropic invites

Song He (Institute of Theoretical Physics, Chinese Academy of Sciences) leads research at the intersection of scattering amplitudes and modern ML tooling. Reporting from September 2026 describes his group's GPT-6-assisted pipeline reaching eight loops on a related six-gluon MHV track, with humans choosing integral bases, interpreting singularities, and signing the final algebra.

The contrast Anthropic's note draws is sharp but narrow:

Neither row is a verdict on which frontier chat model is "best at math" in the abstract — the same caution explainx.ai applied to unconfirmed GPT-6 Pro / Max tier sightings applies here. Loop amplitude work is harness- and oracle-dependent. A higher loop order with SLAC sign-off is evidence about this workflow, not a universal leaderboard score.

What is fair to say: two serious amplitude groups independently bet that frontier models plus heavy tooling beat pencil-and-paper alone for high loop orders — and both still put expert physicists on the critical path for conventions and acceptance.

Builder angle — long-horizon agentic science runs

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

1. Pick problems with cheap falsifiers

Nine-loop amplitudes fail loudly when collinear limits break. Riemann bounds fail when a Lean check fails. Navier-Stokes claims get messy when the stated theorem is not the historical problem.

Agent loops without falsifiers drift. Agent loops with them can run for days.

2. Checkpoint like you mean it

Anthropic's description matches what loop engineering guides preach for code: turn boundaries, persisted artifacts, rollback on test failure. For science runs, replace unit tests with factorization limits, dimension checks, or numeric spot evaluations.

3. Budget a second implementation

SLAC validation is the science version of two independent CI pipelines. If you only have one algebra environment, you are one silent convention bug away from a viral wrong result.

4. Separate model spend from human sign-off

$1–2k compute does not remove the need for a named validator. Plan reviewer time even when tokens are cheap — especially if you publish.

5. Do not conflate "hard calculation" with "foundational proof"

This result belongs next to incremental amplitude milestones, not next to mass-gap claims. Your internal comms should say "verified calculation", not "solved Yang-Mills." Readers coming from Millennium Prize explainers will thank you.

What people are asking

Is nine loops a world record? In planar N=4 SYM six-gluon MHV, nine loops is an aggressive frontier order — the kind of territory where previously published human calculations trailed by one loop or shared partial results. Treat Anthropic's claim as the latest reported order, pending wider community digestion, not an immutable record stone.

Could I reproduce this on public Claude? Anthropic's note references a research orchestration stack — not necessarily the same model tier as consumer Claude Code. Reproducibility for outsiders likely waits on artifact release and pinned model IDs, the same gap that blocked immediate replication of the Riemann bound.

Does this help my coding agent project? Indirectly, yes. The harness patterns — checkpointed tool loops, external verification, spend caps with restart — mirror production agent design. It does not give you a physics integrator out of the box.

Where does this leave GPT-6 for math? Song He's eight-loop GPT-6-assisted track shows both major labs' models can sit inside expert pipelines. Unconfirmed UI tiers like GPT-6 Pro sightings are irrelevant to whether research harnesses can push loop order — until OpenAI ships stable research APIs with published evals on amplitude tasks.

Honest limitations

• Not peer-reviewed journal publication as of September 26, 2026 — research note plus SLAC validation, not a Physical Review Letter acceptance.
• Planar N=4 SYM ≠ QCD — phenomenology pipelines still need separate technology.
• Mass gap untouched — do not merge this story with Millennium Prize Yang-Mills headlines.
• Cost figure is Anthropic's estimate — your rerun may differ with public pricing and retry storms.
• Song He comparison relies on contemporaneous reporting — read both primary notes side by side before declaring a "winner."

Related reading on explainx.ai

• Millennium Prize Problems and AI — fact-check hub (updated for amplitude vs mass gap)
• 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
• Loop engineering for coding agents (Claude Code patterns)
• GPT-6 Pro sightings vs confirmed math harness results
• Will AI replace mathematicians? IEEE Big Mathematics take
• Claude Code loops — /goal, schedules, and turn boundaries
• Fields Medalists' declaration on AI labs and math alignment

Amplitude conventions, loop-order claims, and validation details reflect Anthropic's September 2026 research note and contemporaneous SLAC validation reporting as of publication date. Loop records and community acceptance may evolve as independent groups reproduce the export bundle.