Claude designs verified quantum circuits that beat published designs, Anthropic says

What Anthropic did

Anthropic says it gave Claude plain-language specifications for eight quantum circuits and an automated evaluator that checks and scores each design — and that Claude, working largely without human input, produced circuits that improve on the best published ones, in some cases by an order of magnitude or more. A full paper is available on arXiv.

The project was led by Andi Gu as part of Anthropic's STEM Fellows Program, with mentorship from Alek Kemeny, a quantum physicist at Anthropic, in collaboration with Susanne Yelin of Harvard University and Mikhail Lukin of Harvard University and QuEra Computing.

How the setup worked

According to Anthropic, the eight tasks were computational primitives — the basic routines larger quantum algorithms are built from — spanning chemistry, optimization, quantum simulation, number theory, and numerical computing.

Three properties made the task a good fit for a language model, Anthropic argues. Quantum circuits can be expressed as code, so Claude's programming abilities apply directly. The class of algorithms targeted can be tested for correctness efficiently, because each input produces one definite output that an ordinary computer can verify. And there is a single objective score to minimise: the number of qubits used multiplied by the number of gates, a proxy for how long the circuit would take to run.

Claude ran in a closed loop: it proposed a design in Python, the external evaluator checked correctness and returned a score, and Claude used that feedback to improve its next attempt. Anthropic says the model it identifies as Claude Mythos 5 handled the loop with no human input between rounds.

The results

Across the eight tasks, Claude's designs used both fewer qubits and fewer gates than the best previously published circuits, Anthropic reports.

Notably, the company attributes the gains to new structural and mathematical insights rather than fine-tuning of existing designs. Compiling a circuit for simulating an iron-sulfide molecule, Claude noticed that the molecule's symmetries and repeating structure meant some results could be computed once and reused instead of recalculated for every part of the system. In other cases, it found that published circuits contained costly components needed only to handle extremely rare inputs, and redesigned them without those components.

Two discoveries are highlighted in the post: a new way to perform modular inversion, the most expensive step in elliptic-curve calculations, and a redesign that makes the costliest step of a leading quantum optimization algorithm about twelve times cheaper.

Why circuit design is hard

The post also explains what makes quantum circuit design different from ordinary programming. A quantum circuit cannot make if-then decisions, because it runs on a superposition of many inputs at once, so it must include the operations for every path every time it runs. And it cannot discard intermediate information: every step must be reversible, so leftover scratch work has to be uncomputed by running the computation that created it in reverse, at the cost of additional gates.

Why it matters

Today's quantum computers have only hundreds to thousands of qubits, and those qubits are scarce and error-prone, so both qubit count and computation time are precious resources. Circuit-level efficiency has historically delivered enormous savings here: Anthropic notes that the estimated cost of computing the ground-state energy of FeMoco, an iron-sulfur cluster relevant to industrial catalysis, fell from roughly 50 trillion Toffoli gates in a 2017 estimate to about 340 million after a series of algorithmic advances — a roughly 150,000-fold reduction achieved by redesigning circuits rather than improving hardware.

If AI systems can automate further reductions of that kind, the resource requirements of useful quantum applications shrink accordingly, and Anthropic suggests practical quantum computers could arrive sooner than previously anticipated. The result also points beyond quantum computing: Anthropic had already observed Claude optimizing classical algorithms, and this work suggests the same capability extends to hardware-adjacent research problems — positioning AI as a practical collaborator in scientific and engineering design rather than just a coding assistant.