Claude found cheaper quantum circuits. The scoreboard is the story.

Anthropic published a science post on October 16 describing how Claude Mythos 5 designed quantum circuits that beat the best published versions, and the paper underneath is worth reading: arXiv:2610.22367, submitted October 15, 46 pages, by Andi Gu, Alek Kemeny, Susanne Yelin and Mikhail Lukin. Across eight circuit-design tasks, in seven the model's designs reduced both Toffoli gate counts and qubit counts against the previous state of the art, sometimes by an order of magnitude or more. The eighth produced something different and arguably more interesting: a new reversible architecture for modular inversion. What makes it unusual is not the model but the measuring instrument.

The evaluator is the experiment

Claude never touched a quantum computer. For each task it got a plain-language specification and a program that checked and scored its answer, then iterated: propose a circuit, receive a score, revise, with no human between rounds. The score was the number of qubits multiplied by the number of gates.

That is a thin instrument, which is exactly why the claim can be checked. Quantum circuits are programs, so a model can write, run and revise them as ordinary code. The class of algorithm Anthropic targeted has a property most AI research claims lack: each input produces one definite output, which an ordinary computer can verify. Correctness is decidable, and cost is a count of objects, not a wall-clock measurement. A third party with a laptop can audit the arithmetic without any quantum hardware. Whether the comparison set was assembled fairly is a separate question, and a harder one.

Why circuits are expensive

Two constraints make this domain hostile to ordinary programming. A circuit cannot branch: it runs on a superposition of inputs, so it has to contain the operations for every path on every run. And it cannot discard information, because every step must be reversible. Scratch work is uncomputed by running its creation backwards, at the cost of gates.

The result is a space-and-time trade-off with no free corner. Uncompute intermediates early and you free qubits, but you may have to recompute them later; retain them and you pay in qubits.

The divsteps result

Point addition on elliptic curves is one of the field's standard benchmarks, because most of its cost sits in a single step, modular inversion. Claude ran that task for 53 hours with what Anthropic calls minimal human input. Divsteps, the classical method, had never produced a practical reversible circuit: each round's branch information is scratch work a classical program throws away, while a quantum version must keep the whole history, costing an enormous number of qubits.

Claude's move was to argue the record need not be stored at all, because the numbers themselves could encode it if a decoding rule recovered it. Early candidate rules worked on small test cases and failed at the full 256-bit size. It then proved algebraic identities about the iteration, derived a closed-form decoding rule, checked it in large-scale simulation, and formalized the claim in Lean. One design in the resulting family uses 12% more qubits than the smallest previous design and 17 times fewer gates.

Reading the data instead of storing it

The second example is a Reed-Solomon decoder inside decoded quantum interferometry, one of the stronger candidates for a superpolynomial speedup on optimization problems. The previous circuit used the extended Euclidean algorithm, whose step count depends on its input, and a quantum circuit must run one fixed sequence of gates. So the old design stretched the work over more rounds than any input needed and ran every operation regardless.

Claude switched to Berlekamp-Massey decoding, which performs the same kind of step every round. Reversibility then demanded storage, and the worst case, an exception in 35 rounds, would have cost 665 qubits. Claude wrote a fast simulator and ran 403 million sample decodings to measure how often exceptions actually occur. Almost every run needed at most one saved value, so the circuit reserves 19 qubits and fails roughly once in 10,000 runs, inside the error rate the task allows. Anthropic says the decoder's gate count falls by a factor of 12 against Google's previous construction.

What the multiplier does not mean

The 12x is a gate count against one prior circuit, not a runtime, and qubits times gates is a proxy for runtime rather than a measurement of it. Nothing here ran on quantum hardware; the circuits were checked and priced classically. The "best previously published" baseline is a literature comparison Anthropic compiled, and the eight tasks were chosen by the authors rather than sampled from a shared benchmark, so the headline "seven of eight" describes a set the lab assembled.

The no-human-input claim also has an edge worth stating. The loop was autonomous, but the run included occasional prompts to broaden the search, plus a standing requirement to develop several qualitatively different approaches in parallel. Anthropic presents both as fixes: it says Claude was conservative by nature, second-guessing improvements and preferring well-known methods, and that these interventions pushed it toward the divsteps result.

FeMoco, and the comparison that flatters

One number carries the stakes. In 2017, computing the ground-state energy of FeMoco, the iron-sulfur cluster at the heart of nitrogenase, was estimated to need about 50 trillion Toffoli gates. Algorithmic work has brought that to about 340 million, a roughly 150,000-fold reduction, corresponding to a few hours on a future machine with about three million physical qubits. Both estimates describe the same 54-orbital model, per the post's footnote citing Low et al. 2025.

Anthropic then contrasts eight years of multi-group algorithmic work with a single 53-hour Claude run. The contrast is rhetorically effective and not like-for-like: those eight years bought a specific chemistry estimate, while the 53 hours bought a circuit primitive. The inference that useful machines arrive sooner is Anthropic's, and it stays Anthropic's until someone reproduces the gate counts.

A pattern, not yet a proof

This is the lab's third science-domain result in about a month, after the biomolecular modeling kernels in September and an enzyme-design project, and the cadence looks like a method rather than a coincidence: pick a domain where a design can be written as code, checked for correctness, and scored by cost. Anthropic's conclusion: "Rather than designing each component on its own, AI could optimize all of them jointly, from the algorithm down to the hardware, to build an entire error-corrected quantum computer tailored to a specific problem."

That is a research program, not a result. Evaluator-guided search is only as good as its evaluator: a model very good at one scoring function tells you little about the parts nobody has written a checker for. What would settle the stronger claim is independent replication of the gate counts from the published constructions, plus a task set the authors did not choose. Until then the defensible version is narrow. An LLM given a checkable score found circuit designs that an ordinary computer can confirm are cheaper, and that is a real advance in a field that measures progress in months of expert time.