Anthropic Says Claude Designed More Efficient Quantum Circuits

Anthropic published on 16 October 2026 research claiming that Claude Mythos 5, working in a closed loop with an automated evaluator and no human input between rounds, found quantum-circuit designs that used both fewer qubits and fewer gates than the best previously published algorithms — “in some cases by an order of magnitude or more.” A companion paper is on arXiv.

These are Anthropic’s own research claims. Results are self-reported and not independently replicated. Do not treat them as third-party verified breakthroughs, and do not read the elliptic-curve work as a claim that Claude (or near-term quantum hardware) can defeat elliptic-curve cryptography or Bitcoin.

How the loop worked

Anthropic says it gave Claude plain-language specifications for eight computational primitives drawn from chemistry, optimization, quantum simulation, number theory and numerical computing, plus an automated evaluator that checks correctness and scores designs on qubits × gates. Claude proposed implementations in Python, received scores, and revised — a closed loop illustrated in the post.

The company frames the point: today’s machines have only hundreds to thousands of qubits, noise piles up, so shrinking circuits matters as much as better hardware. As context (Anthropic citing prior literature), it notes the FeMoco ground-state estimate fell from about 50 trillion Toffoli gates (2017) to about 340 million — a roughly 150,000-fold reduction — mostly from algorithmic redesign, not hardware leaps.

Example 1: elliptic-curve point addition (modular inversion)

Anthropic asked Claude to optimise elliptic-curve point addition, a standard benchmark because it dominates costs in quantum algorithms for the elliptic-curve discrete logarithm. After 53 hours with minimal human input (initial brief plus occasional prompts to think more broadly), Anthropic says Claude produced a new reversible approach to divsteps-based modular inversion: the first reversible implementation of divsteps with no stored branch history, checked by simulation and formalised in Lean.

One design in the resulting family, Anthropic reports, uses 12% more qubits but 17× fewer gates than the smallest previous design, expanding the Pareto frontier between space and time. That is a circuit-resource claim about published constructions — not a statement that cryptography is broken, that Bitcoin is at risk from this paper, or that a cryptanalytically relevant machine exists.

Example 2: DQI Reed–Solomon decoder

On the costliest step of decoded quantum interferometry (DQI) — a Reed–Solomon decoder — Anthropic says Claude switched foundations (Berlekamp–Massey instead of the extended Euclidean algorithm), reconstructed most temporary values instead of storing them, and used a high-speed simulator over 403 million sample decodings to size memory for rare exceptions. Reserved memory for one saved value: 19 qubits versus a 665-qubit worst-case reserve. About one run in 10,000 fails, within the allowed error rate, Anthropic says. Gates fell by a factor of 12 relative to Google’s previous construction.

The “conservative by nature” caveat

Anthropic is explicit about limitations. In early experiments, Claude was “conservative by nature.” It repeatedly second-guessed apparent improvements and sometimes double- and triple-checked already well-supported results. That caution was sometimes useful (for example, writing Lean proofs for the divsteps design), but it also showed a tendency to prefer existing, well-known approaches over novel ones.

Researchers mitigated that by requiring several qualitatively different approaches in parallel and by occasionally prompting Claude to broaden its search. The elliptic-curve result, Anthropic says, came from exploring a method that had not been practical before — not from shaving gates off a pre-existing design.

People and framing

The work was led by Andi Gu (Anthropic STEM Fellows Program), mentored by Alek Kemeny, with Susanne Yelin (Harvard) and Mikhail Lukin (Harvard and QuEra Computing). Anthropic argues that faster circuit design could bring useful applications closer and that “practical quantum computers could arrive sooner than previously anticipated” — an aspirational company line, not a dated forecast of cryptographically relevant machines.

Why it matters — carefully

If the self-reported improvements hold under independent scrutiny, AI-assisted circuit compilation becomes another lever beside hardware and error correction. TSN has covered adjacent AI-for-quantum engineering stories, including IBM’s agent-sped quantum decoder chip design, Google Quantum AI’s magic-state cultivation on Willow, and NVIDIA ISING. For wallet-key risk context that is not this paper’s claim, see Quantum and wallet keys.

What’s still unknown

• Independent replication of the eight primitives and the 17× / 12× resource claims
• How much human prompting to “broaden search” was required beyond the closed loop
• Whether Lean formalisations cover every claimed identity end-to-end

What this means

• Anthropic: Claude Mythos 5 found smaller/faster circuits than published baselines on eight primitives — company claims.
• Highlights: reversible divsteps (Lean-checked); DQI decoder ~12× fewer gates vs Google prior (Anthropic).
• Model was “conservative by nature”; researchers prompted broader search.
• Not independently replicated; not a claim that elliptic-curve crypto or Bitcoin is defeated.

Sources

1. Anthropic, “Designing more efficient quantum circuits with Claude”, 16 October 2026
2. arXiv:2610.22367 (companion paper linked from Anthropic)

Related stories

• IBM Research Says Open AI Agents Sped Up a Quantum Decoder Chip Design
• Google Quantum AI Shows “Magic State Cultivation” Working on Its Willow Chip
• NVIDIA ISING: The Open-Source AI Bridge to Practical Quantum Computing
• Quantum and wallet keys: Europol says the risk is exposed keys