Anthropic said on October 16, 2026 that Claude, working largely without human input, designed quantum circuits that need fewer qubits and fewer gates than the best previously published designs across a set of eight standard tasks. The project was led by Andi Gu through Anthropic's STEM Fellows Program, with Anthropic quantum physicist Alek Kemeny and Harvard's Susanne Yelin and Mikhail Lukin, who is also affiliated with QuEra Computing. A 46-page paper, "Discovering state-of-the-art quantum circuits with large language models", was posted to arXiv on October 15 (arXiv:2610.22367).
The setup is a closed loop. Claude Mythos 5 received a plain-language specification for each circuit and wrote candidate implementations in Python; an independent evaluator program checked each one for correctness and scored it on qubits multiplied by gates, a proxy for runtime; Claude used the score to try again, with no human input between rounds. The eight primitives cover numerical computation, chemistry, optimization, quantum simulation and number theory. According to the paper's abstract, the discovered circuits significantly reduce both Toffoli gate and qubit counts relative to the state of the art in seven of the eight tasks, and in the eighth the search produced a new reversible architecture for prime-field modular inversion.
Quantum circuits are awkward to optimize because they cannot branch the way ordinary programs do and cannot simply discard intermediate results. Every step must be reversible, so scratch work has to be uncomputed at the cost of extra gates, or kept at the cost of extra qubits. That trade-off is what Claude was asked to push against.
The headline example is elliptic-curve point addition, a standard benchmark because it dominates the cost of quantum attacks on the elliptic-curve discrete logarithm problem. Its most expensive step is modular inversion. A classical method called divsteps is efficient on ordinary computers but had not produced a practical quantum circuit, since a reversible version must store the full history of which branch was taken. After a 53-hour run with minimal prompting, Claude proposed recovering that branch information from the pair of numbers itself, proved new algebraic identities to derive an exact decoding rule, checked it with large simulations and formalized it in Lean. Anthropic calls the result the first reversible divsteps implementation with no stored branch history. By choosing when to store rather than decode, Claude then produced a family of designs; one uses 12% more qubits but 17 times fewer gates than the smallest previous design.
A second result targets decoded quantum interferometry (DQI), a candidate route to a superpolynomial quantum speedup on optimization problems, where Reed-Solomon decoding is the costliest step. Claude replaced the extended Euclidean algorithm with Berlekamp-Massey decoding, reconstructed most temporary values instead of storing them, and ran 403 million simulated decodings to size a single 19-qubit reserve rather than a 665-qubit worst case, accepting roughly one failure in 10,000 runs, within the task's tolerance. Anthropic says the decoder needs 12 times fewer gates than Google's previous construction.
Anthropic frames the work against the slow history of resource estimates: the gate count for simulating the FeMoco enzyme cluster fell from about 50 trillion Toffoli gates in 2017 to about 340 million through eight years of work by many groups, while the elliptic-curve design came from one 53-hour run. The company suggests the same evaluator-driven approach could extend to error-correcting codes, decoders and hardware control.
There are clear limits. Anthropic reports that Claude was conservative early on, repeatedly second-guessing improvements and favouring familiar methods, so the researchers required several different approaches in parallel and occasionally prompted it to search more broadly. The results are resource estimates for future fault-tolerant machines, not runs on quantum hardware, and the idea that useful quantum computing could arrive sooner is Anthropic's projection. The paper is a preprint, and no independent replication was found at publication.