Claude Just Broke a Physics Record That Had Stood Since 2023

A frontier AI model just did something no human physicist has managed: it pushed a notoriously brutal particle physics calculation one loop past the previous world record, working almost entirely on its own for several days straight. Anthropic says its Claude Opus 5-class model autonomously computed a nine-loop scattering amplitude in a six-particle process, beating an eight-loop record that had stood since roughly 2023.

If that sentence means nothing to you, stick around. The short version: this is one of the hardest categories of calculation in theoretical physics, the kind where each additional step multiplies the difficulty by an order of magnitude, and an AI system just went further into that difficulty than any person or research group ever has.

What Actually Got Calculated

The object in question is called a scattering amplitude. When particles collide, physicists don’t get a single clean answer about what happens next — they get probabilities, and scattering amplitudes are the mathematical machinery that produces them. You build these amplitudes as what’s called a perturbative expansion, essentially a series of increasingly precise correction terms. Each correction is a “loop,” and each loop you add makes the math dramatically messier.

To put the difficulty in perspective, most scattering amplitude calculations in physics never get past two loops. A short list of especially punishing problems has reached three. The calculation of the electron’s anomalous magnetic moment — arguably the most precisely tested prediction in all of physics — took human researchers to five loops, and that project is treated as a landmark of theoretical physics.

Claude went to nine.

The Specific Record

The calculation Claude tackled involves the six-particle amplitude in planar N=4 super Yang-Mills theory, a highly symmetric quantum field theory that physicists use as a testbed rather than a description of the real world. It’s sometimes called a “toy” theory, but that undersells it — its symmetries are precisely what make it possible to push loop calculations to depths that would be hopeless in real-world quantum chromodynamics, the theory that actually governs quarks and gluons. It’s where new computational techniques get proven out before anyone tries to apply them to messier, more physically realistic problems.

The previous benchmark for this exact six-particle amplitude — eight loops — belonged to Lance Dixon of SLAC National Accelerator Laboratory and his collaborators, a result that had held since around 2023. Claude’s nine-loop result doesn’t just match that mark, it extends it, in a field where extending it by even a single loop can represent months of specialized human effort.

How Claude Actually Did It

What makes this more than a brute-force stunt is the method. Claude didn’t throw raw computing power at the problem until an answer fell out. It used two complementary bootstrap techniques that are themselves established tools in this corner of theoretical physics.

• Direct bootstrap: Rather than deriving the amplitude term by term, this approach starts from candidate mathematical functions and narrows them down by imposing every known physical and mathematical constraint the true answer has to satisfy — symmetry properties, consistency conditions, known limiting behavior — until only one function survives.
• Form-factor route: This path calculates a related but more manageable quantity first, then translates that result back into the full amplitude using a mathematical relationship known as antipodal duality.

Running both approaches and cross-checking them against each other is standard practice for physicists doing this kind of work. Claude reportedly picked and executed both strategies on its own, without a human directing the mathematical approach step by step.

Loop Records at a Glance

Days of Autonomous Work, Minimal Human Input

Perhaps the most striking detail isn’t the loop count — it’s the duration and independence of the run. Claude worked on this problem for multiple days. Researchers weren’t silent the entire time; they periodically sent “continue” prompts to keep the session going. But there was no mid-run scientific correction, no human stepping in to fix a wrong turn in the math or redirect the method. The reasoning, the choice of technique, and the execution were the model’s own, sustained over a stretch of time long enough for a human team to lose momentum, make errors, or simply run out of patience.

That’s a different kind of achievement than getting a single correct answer to a hard prompt. Staying coherent and correct across a multi-day unsupervised research process is its own test, separate from raw calculating ability, and it’s the part of this story that should catch the attention of people well outside particle physics.

What It Cost

The financial numbers here are almost as notable as the physics. Anthropic put the total cost of the achievement at roughly $1,000 to $2,000, dominated by the price of running the model itself. The numerical bootstrap portion alone cost around $100 in compute, which researchers compared to running 96 CPUs continuously for a full week. That’s a rounding error next to what a human research team would spend in salaried time chasing the same result, if they could chase it down at all.

The Necessary Caveat

None of this means Claude invented a new law of physics or discovered a technique nobody had thought of before. Experts studying the result are clear on that point: this is a case of an AI system executing already-established scientific methods — bootstrap techniques physicists have used for years — extremely well, over an unusually long and unsupervised stretch, and producing a result other physicists can independently check. It’s execution at a new scale, not discovery of something new.

That distinction matters, but it shouldn’t undersell what happened. Plenty of skilled human teams understand these bootstrap methods in principle and still can’t sustain the kind of long, error-free, multi-day grind this problem demanded. The ability to do that without going off the rails is exactly the capability that’s been missing from AI systems tackling frontier science, and it’s the part of this story most likely to generalize to other hard problems.

What This Means

Zoom out and this result lands in a busy week for Anthropic. The company has said Claude now contributes to roughly 26% of its own internal AI research work, and reports have the company moving toward a large IPO. A physics record isn’t proof of any of that momentum by itself, but it’s a concrete, independently verifiable data point at a moment when a lot of AI capability claims are hard to pin down.

The real takeaway isn’t that AI is now better than physicists at physics. It’s that a model can now sit with one exceptionally hard, narrow scientific problem for days, choose its own methods, avoid the kind of drift or error that would sink the result, and hand back an answer that specialists can check and trust. That’s a narrower claim than “AI does science now,” but it’s also a more useful one — and it’s the kind of quiet capability jump that tends to matter more in a year than it does the week it’s announced.