Claude autonomously designs novel quantum circuits beating published benchmarks
Designing more efficient quantum circuits with Claude (external)
Anthropic reports an AI system cut the gate count of two benchmark quantum subroutines (one by 17x, one by 12x) in autonomous closed-loop runs, so the resource estimates that gate fault-tolerant applications may fall faster than hardware timelines assume.
- Setup: 'Claude Mythos 5' ran a closed loop proposing a circuit, getting a score from an automated evaluator (objective = qubits x gates as a runtime proxy), and revising, with no human input between rounds; it was tested on 8 computational primitives across chemistry, optimization, physics simulation, number theory, and numerical computing. All results are Anthropic self-reported; full paper is on arXiv.
- Modular inversion (elliptic-curve point addition): Claude made divsteps reversible without storing the full branch-information history by deriving a decoding formula (new algebraic identities, checked via large-scale simulation and formalized in Lean). One resulting circuit uses 12% more qubits but 17x fewer gates than the smallest prior design, and the family extends the qubit/runtime Pareto frontier. This came from a single 53-hour run.
- DQI / Reed-Solomon decoding: Claude replaced the extended Euclidean algorithm with Berlekamp-Massey (uniform rounds suit a fixed gate sequence), reconstructed temporary values instead of storing them, and ran 403M sample decodings to size memory empirically, cutting worst-case 665 qubits to 19 by accepting a ~1-in-10,000 failure rate. Net: 12x fewer decoder gates than Google Quantum AI's prior construction.
Full summary
Anthropic describes using a model it calls Claude Mythos 5 to redesign quantum circuits autonomously, running a closed loop in which Claude proposes a circuit, an automated evaluator scores it on qubits times gates, and Claude revises without human input between rounds. Across eight computational primitives spanning chemistry, optimization, quantum simulation, number theory, and numerical computing, the company says Claude found designs beating the best published circuits, some by an order of magnitude or more. Two cases are detailed: a reversible divsteps-based modular inversion for elliptic-curve arithmetic that avoids storing branch history by decoding it from the number pairs (one circuit uses 12% more qubits but 17x fewer gates, verified in Lean, from a single 53-hour run), and a Reed-Solomon decoder for Decoded Quantum Interferometry that switches to Berlekamp-Massey, reconstructs temporary values, and uses empirical sampling (403M decodings) to shrink worst-case memory from 665 to 19 qubits at a ~1-in-10,000 failure rate, for 12x fewer gates than Google's prior design. Anthropic also reports a limitation: Claude defaulted to conservative, known approaches and needed prompts to pursue parallel, more ambitious ideas. All figures are self-reported, with the full paper posted to arXiv; the work was led by Andi Gu via Anthropic's STEM Fellows Program with Harvard and QuEra collaborators.