A closed loop, a score, and eight hard routines

Quantum algorithms spend most of their cost on a small set of building-block routines. Make one routine cheaper in gates or qubits and every algorithm built on it gets cheaper. That is the layer Anthropic aimed Claude at. Each task arrived as a plain-language specification. An evaluator program checked whether a proposed circuit was correct and scored it on qubits multiplied by gates. Claude wrote designs as code, took the score, and tried again, with no human input between rounds inside the loop.

The setup matters more than the slogan attached to it — but it should not be oversold either. Anthropic did not point the model at circuit design in general. It selected a class of computations whose implementations compute defined functions, where each input gives one definite output an ordinary computer can check efficiently. Tasks of that shape are particularly suitable for automated evaluation: an answer can be verified mechanically and scored against an objective proxy. That is exactly the shape of problem a model can iterate against without a person grading each attempt. The paper's claim is that what came out of the loop was not fine-tuning. In the two cases Anthropic details, the model changed the underlying method.

Where the 17x comes from, and what it costs

The first case is modular inversion, the expensive step inside elliptic-curve point addition. A classical method for it, divsteps, had never yielded a practical quantum circuit, because a quantum circuit cannot throw away its scratch work: the record of which operation ran at each step — the branch information — has to be kept so the computation can be reversed, and keeping it costs a large number of extra qubits. Anthropic gave Claude a one-paragraph brief, and after about 53 hours of model-driven exploration, with minimal human input limited to that brief and occasional prompts to think more broadly, the model produced a version that stores no such record at all.

Per Anthropic's account, Claude derived a rule for reconstructing the branch information from the numerical state itself: it proved a series of algebraic identities about the divsteps iteration, derived a precise decoding formula from them, checked the formula in large-scale numerical simulations, and formalized the central claim in Lean. That formalization covers the central mathematical claim; it is not an end-to-end formal proof of the whole circuit implementation. The result is described as the first reversible implementation of divsteps with no stored branch history.

The headline number needs its other half. One design in the resulting family uses 17 times fewer Toffoli gates than the smallest previously published design, and 12% more qubits. Two qualifications belong with it. First, the comparison concerns the paper's reported logical gate counts — the Toffoli gate is one particular, expensive class of reversible gate, and counting it is a standard way to estimate a computation's cost — not measured execution time on physical quantum hardware. Second, a gate-count reduction is not necessarily an equal reduction in wall-clock time or overall physical cost, because different operations carry different costs. Gates and qubits are both scarce, and they trade against each other. A 17x gate reduction bought with a 12% qubit increase is a striking theoretical trade, and its practical benefit depends on the target machine and architecture, and on how those gates translate into physical resources. It is also one design point, for one routine, against 2026 comparison papers. It is not a 17x improvement in quantum computing.

The second case makes the same point more quietly. For the decoder inside decoded quantum interferometry, a leading optimization algorithm, Claude replaced the previous construction's method — the extended Euclidean algorithm — with Berlekamp–Massey decoding, whose steps are uniform across inputs. Anthropic reports that Claude then ran about 403 million simulated sample decodings to estimate how much temporary storage the redesigned decoder actually needed: a worst-case reserve of 665 qubits could be replaced by 19 qubits set aside for a single saved value. Those simulations support an empirical estimate under the tested distribution; they do not prove the smaller reserve suffices in every case, and about one run in 10,000 needs more than the reserve and fails — which Anthropic says sits well inside the error rate the task allows. The gate count fell by a factor of 12. No new physics. Just a budget set by measurement instead of by worst case.

How much of this to bank

Three discounts apply, and Anthropic's own post supplies all three. First, the model is Mythos 5, an internal research system. Nothing here can be tried by a reader, a customer, or a rival lab with the same tool. Second, the paper is a preprint. The designs are public, and the preprint supplies mathematical descriptions, named comparisons and resource estimates that outsiders can examine and re-score — a tier above an internal benchmark announcement, but not a fully reproducible experimental package, and peer review has not happened yet. Third, the circuits are designs. Today's machines hold hundreds to thousands of noisy qubits; these routines are priced for future fault-tolerant hardware.

That third discount needs stating carefully. A more efficient circuit could reduce the logical resources a future fault-tolerant computation requires, and Anthropic argues that improvements of this kind could bring useful applications within reach sooner. What this work does not do is establish any new date for that hardware. The post's own framing example cuts both ways: the field spent eight years and many research groups cutting one famous chemistry estimate — the FeMoco simulation — by a factor of roughly 150,000, and the machine for it still does not exist.

What survives the discounts is the method. An evaluator-guided search did not just shave constants off existing circuits; in two documented cases it swapped in a different method and then did the unglamorous work — identities, simulations, a 403-million-run measurement — needed to defend the swap. That is a new way to staff a scarce specialty. Quantum circuit designers are rare. A process that produces checkable designs for them to attack is useful even if every number in this paper is later revised downward.

Sources

  1. [1] Anthropic (primary), "Designing more efficient quantum circuits with Claude", research post, 16 October 2026Read source
  2. [2] Gu, Kemeny, Yelin and Lukin (primary), "Discovering state-of-the-art quantum circuits with large language models", arXiv:2610.22367, submitted 15 October 2026Read source
  3. [3] Unite.AI, "Anthropic Says Claude Beat Best Published Quantum Circuit Designs", 16 October 2026 (pickup confirmation only; no claim in this article rests on it alone)Read source