Anthropic: Claude Mythos 5, given plain-language specs and an automated scorer, designed quantum circuits that beat the best published ones on 7 of 8 tasks, cutting combined cost by up to ~39x

Why it matters

【Key points】

Anthropic published "Designing more efficient quantum circuits with Claude" on Oct 16 at 17:29 UTC, with a paper (arXiv 2610.22367) by Andi Gu (Anthropic STEM Fellows Program), Anthropic quantum physicist Alek Kemeny, and Harvard's Susanne Yelin and Mikhail Lukin.

Claude Mythos 5 got only a plain-language task spec and an automated evaluator that checks correctness and scores qubits × Toffoli gates. It then iterated inside the open-source OpenEvolve search loop. Across eight primitives from chemistry, optimization, quantum simulation and number theory, seven designs cut both qubits and gates. The combined cost fell by up to about 39x (fixed-point Gaussian), 32x (HHL rotation angle) and 12x (Reed–Solomon DQI decoder). For elliptic-curve point addition, a 53-hour run produced a reversible divsteps modular inversion that stores no branch history, with a proof checked in Lean; one variant uses 12% more qubits but 17x fewer gates than the smallest prior design.

【Verifiable】

Anthropic post: https://www.anthropic.com/research/designing-efficient-quantum-circuits

Paper: https://arxiv.org/abs/2610.22367

Anthropic on X: https://x.com/AnthropicAI/status/2111152183240511640

【Our read】

1. Another hard case of a model plus a machine-checkable scorer doing research: the output is checkable circuits and cost tables, not a demo.

2. Know the limits: baselines are the best published circuits, and humans set the specs, scoring and failure budgets, plus occasional prompts to think more broadly. Anthropic says early Claude was conservative and over-checked.

3. Practical quantum computing is still far off; cheaper circuits are one piece, while error correction and hardware dominate. Watch for independent replication by quantum groups.