Anthropic Says Claude Cut Quantum Circuit Gates 17x on One Task

Anthropic said Oct. 16 that Claude designed quantum circuits using fewer qubits and fewer gates than the best published designs on all eight tasks it was given, including one elliptic-curve design with 17 times fewer gates. The company's research post points to a paper, arXiv:2610.22367, submitted Oct. 15.

The work was led by Andi Gu in Anthropic's STEM Fellows Program and mentored by Alek Kemeny, a quantum physicist at Anthropic. Susanne Yelin of Harvard and Mikhail Lukin, of Harvard and QuEra Computing, are co-authors on the paper. Both the post and the paper are the authors' own account.

How the loop worked

Claude received plain-language specifications for eight circuit primitives spanning chemistry, optimization, physics simulation, number theory and numerical computing. An automated evaluator checked each proposed circuit for correctness and scored it by qubits multiplied by gates, a proxy for runtime. Claude then revised its designs from that feedback with no human input between rounds.

Humans still shaped the run. Anthropic said the team supplied the initial brief and occasional nudges, and that Claude was conservative, repeatedly second-guessing itself and favouring well-known approaches. The team required several distinct approaches in parallel to counter that.

The two results with numbers

The elliptic-curve task, point addition, took Claude about 53 hours. It built a reversible version of divsteps, a modular-inversion method, that reconstructs branch information from the numbers instead of storing it. One resulting design uses 12 percent more qubits than the smallest earlier design and 17 times fewer gates. Verification used large numerical simulations at 256 bits and a Lean formalization.

The second result is a Reed-Solomon decoder for the optimal polynomial intersection problem within DQI. Claude swapped the extended Euclidean algorithm for Berlekamp-Massey and recomputed temporary values instead of keeping them. The decoder reserves 19 qubits, against 665 in the worst case, and uses about 12 times fewer gates than Google Quantum AI's earlier construction. Across 403 million sample decodings, it failed about once in 10,000 runs.

Task	Reported change
Elliptic-curve point addition	12% more qubits, 17x fewer gates
Reed-Solomon decoder (DQI)	19 qubits vs 665 worst case
Same decoder vs Google Quantum AI	About 12x fewer gates

The arXiv abstract is more guarded than the post. It says seven of the eight tasks produced circuits with reduced Toffoli and qubit counts compared with the state of the art, and that the eighth produced a new architecture for prime-field modular inversion. The abstract gives no percentage reductions.

What it does not show

Anthropic said circuit design is one part of a fault-tolerant quantum computer. Error correction, decoders and hardware control are separate problems. The post also reproduces a history for FeMoco, a chemistry target: the Toffoli estimate fell from about 50 trillion in 2017 to about 340 million in 2025, roughly 150,000-fold. Anthropic credits that to many researchers' algorithmic work, not to Claude.

The numbers also do not say how many qubits an attack on a real elliptic-curve key would need. Neither the post nor the abstract gives that total.

For related reading, see the site's earlier report on Anthropic's NIST curve seed bounty and on Google's magic state cultivation result.

The paper is version 1 on arXiv, 46 pages with 10 figures and six tables, so any revision will carry a new version number.