In 53 hours, Claude reworked the key operators of 8 quantum circuits—on several of them, beating the best published designs from the past few years by another order of magnitude. That is the headline from Anthropic's October 16 study, "Designing more efficient quantum circuits with Claude," with the paper simultaneously posted to arXiv (arXiv:2610.22367).
Why quantum circuits are hard
A quantum computer cannot branch like a classical one: every input runs through the same fixed instruction sequence, so every branch must be executed in every round. Each step must be reversible; intermediate results cannot simply be discarded, only uncomputed at additional gate cost. Quantum circuit design is therefore a constant trade between qubit count (space) and gate count (time). A landmark case: the Toffoli gate count for FeMoco simulation dropped from 50 trillion in 2017 to 340 million in 2025, roughly 150,000x, almost entirely from circuit redesign rather than hardware improvements.
The question Anthropic asked: what happens when you hand the same "design, evaluate, iterate" loop to an LLM?
53 hours, 8 circuits
The team gave Claude Mythos 5 a plain-language math brief (for example, the rule for elliptic curve point addition over a prime modulus) plus an automated evaluator that checked each candidate circuit for correctness and scored it on qubit count multiplied by gate count. Claude submitted implementations, received feedback, revised, and submitted again, with almost no human intervention. Across 8 circuits spanning chemistry, optimization, quantum simulation, number theory, and numerical computing, Claude's designs beat the best published implementations on both qubit count and gate count. Two stand out.
divsteps: eliminating branch history storage
The most expensive step in elliptic curve point addition is modular inversion. divsteps is efficient on classical computers but runs into the reversibility wall on quantum circuits: which branch was taken at each iteration must be stored in order to uncompute, ballooning qubit cost. Claude's idea: instead of storing branch information, decode it directly from the pair state. Early decoding heuristics worked on small cases and broke at 256-bit sizes. Rather than continue enumerating, Claude pivoted to theory, proved a new set of algebraic identities, derived the correct decoding rule from those identities, validated it at scale with numerical simulations, and formalized the derivation in Lean. The result is the first reversible divsteps implementation that needs no branch history storage. One design in the family uses 12 percent more qubits but cuts gate count to 1/17 of the previous smallest.
Reed-Solomon decoder: 12x fewer gates
The second highlight is the Reed-Solomon decoder, the most expensive step in Decoded Quantum Interferometry (DQI). Google Quantum AI's 2025 circuit used the extended Euclidean algorithm, where different inputs take different paths, so all paths had to be unrolled in the quantum circuit and gate count grew large. Claude switched to Berlekamp-Massey decoding, where every iteration takes the same kind of step and the same number of rounds, a much better fit for quantum's fixed-instruction requirement. But reversibility introduces a new problem: each round's temporary values must either be stored or reconstructed. Claude again chose reconstruction; running 403 million sample decodings showed that in nearly all cases only a single 19-qubit value needs to be saved, instead of the worst-case 665. Algorithm swap, architectural change, and tighter resource budget together cut decoder gate count by 12x.
What this means
This is not another "AI beats a benchmark" story. It is a closed loop in which an AI derived new theory on a problem without a ready template, wrote Lean proofs, ran the resulting circuits, and quantified the speedup. The divsteps case is the one to remember: Claude did not micro-optimize an existing algorithm. It identified the reversibility constraint as the core obstacle, proved a new set of algebraic identities from scratch, and used those identities as the foundation for a circuit redesign—something ordinary automated optimization cannot do.
For quantum hardware roadmaps, the practical implication is more modest than the headline: this work did not claim a universal quantum computer is around the corner. It pushed the cost of a handful of well-studied operators down by another order of magnitude. But Anthropic pointed to the broader possibility: since circuit design, error-correcting codes, error decoders, and hardware control all share the property that designs can be written as code, correctness can be checked automatically, and there is an objective score to optimize, could an AI eventually co-optimize all of them rather than each being tuned separately? This paper is arguably the first quantifiable waypoint on that path.
For readers outside quantum, the more immediate takeaway is about how AI shows up in serious scientific discovery today: less as a coding tool, more as a collaborator that can stay in a closed loop, keep probing a mathematical structure, and arrive at a new result. The Lean proof Claude wrote as part of the divsteps verification is one concrete shape this new role takes.