Anthropic says Claude cut a quantum circuit’s gate count 17-fold

Anthropic says AI-designed circuits could bring useful quantum computers closer. Its October 16 research offers a narrower, concrete result: Claude Mythos 5 improved resource use across eight circuit-design tasks. The accompanying paper distinguishes seven tasks with reductions in both gates and qubits from an eighth with a new architecture and better tradeoffs.

The study covers basic routines used in chemistry, optimization, quantum simulation, number theory and numerical computing. Andi Gu led the work through Anthropic’s STEM Fellows Program, mentored by quantum physicist Alek Kemeny, with collaborators Susanne Yelin and Mikhail Lukin. Their paper was submitted to arXiv on October 15.

A search loop with a checkable answer

Researchers gave Claude plain-language specifications and an automated evaluator. Claude wrote circuit implementations in Python; the evaluator checked correctness and returned a score. The model then revised its design. Anthropic says this proposal-and-feedback cycle ran without human input between rounds.

That setup suited the circuits the team selected: each input produces one definite output that an ordinary computer can check. The score multiplied qubit count by gate count, giving Claude an explicit resource target. Qubits store quantum information; gates are the operations a circuit applies to them. The product served as a proxy for runtime, rather than a measured execution time.

The paper reports lower qubit and Toffoli-gate counts on seven tasks. Toffoli gates are particularly expensive operations, making their count a standard way to estimate computational cost. For the eighth task, the result was a new reversible method for modular inversion—an arithmetic operation central to calculations on elliptic curves—not a blanket reduction in both resources.

Recovering the scratch work instead of storing it

Quantum circuits face a constraint ordinary programs do not: their computational steps must be reversible. Temporary information cannot simply be deleted. Clearing scratch work requires reversing the operations that created it, adding gates. Keeping intermediate results avoids recomputation but consumes qubits, creating a persistent tradeoff between memory and time.

Claude tackled that problem through divsteps, a classical method for modular inversion. A straightforward reversible version must retain a history of the decisions made during its iterations. Claude instead sought to recover those decisions from the numerical values left after each step, avoiding the need to store the full history.

Early decoding rules passed small tests but failed on 256-bit numbers. Anthropic says Claude then derived algebraic identities and a decoding formula, checked it through large numerical simulations, and formalized the claim in Lean. The point-addition effort took 53 hours, with human input limited to the initial brief and occasional prompts to think more broadly.

Recovering the history also takes computation. Claude therefore developed a family of designs that sometimes stores information and sometimes reconstructs it. Changing that threshold produced combinations of memory use and gate count unavailable in earlier published designs, rather than one circuit that wins on every measure.

A second bottleneck, fewer gates

Another result targeted the decoder used by decoded quantum interferometry, a quantum optimization algorithm. Anthropic says Claude reduced the decoder’s gate count by a factor of 12 relative to Google’s earlier construction. The redesign combined a different decoding method, reconstruction of temporary values instead of storage, and a more precise resource budget.

Circuit redesign has already changed estimates of what useful quantum computation might require. Anthropic cites FeMoco, an iron–sulfur cluster important to nitrogen-fixing bacteria: successive algorithmic advances reduced one calculation’s estimated cost from about 50 trillion Toffoli gates to about 340 million. That history explains the stakes of reducing circuit costs. It is context for this study, not a gain attributable to Claude.