Claude Slashes Quantum Circuit Costs, but the Viral 17x Figure Needs Context

Anthropic and Harvard researchers have published results showing that Claude, working largely unsupervised, can rewrite the programs that describe quantum circuits so they consume less of the two resources that matter most: Toffoli gates and logical qubits. Across eight computational primitives drawn from chemistry, optimisation, quantum simulation, number theory and numerical computing, the model beat the best published designs on seven of them. The widely repeated claim of a 17x gate reduction, however, rests on a narrower comparison than the headlines suggest.

• Closed-loop design: Claude Mythos 5 proposed circuits as Python code, an independent evaluator verified correctness and scored each attempt, and the strongest candidates advanced through an evolutionary search framework called OpenEvolve.
• Uneven but real gains: combined Toffoli-qubit cost dropped 39x on a fixed-point Gaussian routine and 32x on HHL angle rotation, yet only 4.1x on the Fe2S2 select step.
• Reading the 17x correctly: the elliptic-curve result accepts 12% more qubits than the smallest prior design in exchange for far fewer gates — it is not a clean sweep against the cheapest design overall.
• Estimates, not hardware: every figure is a resource estimate for future fault-tolerant machines, taken from a 46-page preprint that has not been peer reviewed.

Inside the closed optimisation loop

The starting point was deliberately thin: a hand-written specification and an empty program. In each round, a fresh session of the model proposed architectural changes and code edits, executed them, consulted a classical oracle holding the correct answer for specific inputs, and was free to delegate subtasks to sub-agents. According to the accompanying paper, there was no human intervention between rounds; for the most demanding task, a single run lasted 53 hours, interrupted only by the initial brief and occasional nudges to search more broadly.

The work was led by Andi Gu within Anthropic’s STEM Fellows programme, with mentorship from the company’s quantum physicist Alek Kemeny and collaboration from Susanne Yelin and Mikhail Lukin of Harvard. One circuit’s correctness was additionally formalised with the Lean proof assistant, an unusual level of rigour for machine-generated design.

Where the savings actually land

Anthropic scores each circuit by multiplying Toffoli count by peak qubit count — a proxy for runtime, since Toffoli gates are among the costliest operations on an error-corrected quantum computer and qubits are scarce and fragile. It is that single combined metric, not gate count alone, that improves on seven of the eight tasks.

The clearest single result sits in a decoder for Google’s DQI algorithm, where Reed-Solomon decoding is the expensive step. Claude swapped the extended Euclidean algorithm for Berlekamp-Massey decoding, reconstructed most temporary values instead of storing them, and sized a 19-qubit reserve rather than a 665-qubit worst case by running 403 million simulated decodings and accepting roughly one failure in 10,000 runs. The resulting circuit needs about 440,000 Toffolis, down from 5.1 million, with peak qubits trimmed from 1,938 to 1,837.

The fine print behind the 17x, and why it matters

The celebrated figure comes from 256-bit point addition on an elliptic curve, the core of Shor-style attacks on elliptic-curve cryptography. Claude devised a reversible implementation of the divsteps method that reconstructs branching decisions from intermediate values instead of storing them. Its circuit uses 4 million Toffolis and 935 qubits, against 70 million Toffolis and 838 qubits in the 2026 reference design by Luo and colleagues.

Anthropic states the trade-off plainly: a design with “12% more qubits but 17 times fewer gates”.

What travelled less widely is the second comparison in the paper’s own table. A rival 2026 architecture by Schrottenloher needs 1.9 million Toffolis and 1,446 qubits — half the gates of Claude’s circuit, at the cost of more memory — and wins on the combined metric. That is why the authors frame the result as pushing out the frontier between qubit count and runtime rather than beating the best design outright, and why the task carries no improvement factor in the main table. A smaller inconsistency also surfaces: Anthropic’s blog post claims fewer qubits and fewer gates across all eight tasks, while the paper says seven of eight.

The strategic argument still holds. Quantum algorithms have historically become cheaper through better circuits rather than better hardware: estimates for simulating the FeMoco iron-sulphur cluster fell from roughly 50 trillion Toffolis in 2017 to about 340 million — some 150,000 times — over eight years of work by many groups. A single 53-hour run producing a new arithmetic architecture suggests that curve could steepen.

Anthropic is candid about the limits. Claude was “conservative by nature” early on, favouring familiar approaches and second-guessing its own improvements, so the team ran several strategies in parallel. The metric is a modelling choice, real costs depend on hardware, and today’s machines hold only hundreds to thousands of noisy qubits.

The most valuable contribution is arguably methodological rather than numerical: a model that proposes, a verifier that cannot be flattered, and dozens of hours of unattended search. That pattern extends well beyond quantum circuits — to error-correcting codes, decoders and hardware control. It also sharpens an uncomfortable angle, since the curve in question underpins Bitcoin, even if the authors never frame the work as a threat. No attack arrives tomorrow, but each circuit-level saving narrows the margin that current public-key cryptography is counting on.