Anthropic says Claude beat the best published quantum circuits on eight tasks

A free-running general-purpose model, told what good looks like by a short scoring program, found quantum circuits that use fewer qubits and fewer gates than the best previously published designs. Anthropic has posted the research alongside an arXiv paper.

Anthropic says that it gave Claude Mythos 5 plain-language briefs for eight quantum computational primitives, from chemistry, optimization, physics simulation, number theory and numerical computing, together with an automated evaluator that checked each proposed design for correctness and scored it on the number of qubits and gates it needed. Working mostly on its own, Claude produced designs that improve on the best previously published algorithms across all eight tasks, in some cases by an order of magnitude or more. The work was led by Andi Gu through Anthropic's STEM Fellows Program, mentored by Alek Kemeny, in collaboration with Susanne Yelin at Harvard and Mikhail Lukin at Harvard and QuEra Computing. The paper is on arXiv.

The two headline results

Point addition is the main step in elliptic-curve quantum algorithms, and most of its cost comes from a routine called modular inversion. Anthropic says Claude derived a new reversible algorithm for divsteps, the fastest classical method for modular inversion, by proving a set of new algebraic identities and then formalising the key claim in Lean. One circuit in the family it produced uses 12% more qubits than the smallest previous design, but 17 times fewer gates, with Anthropic crediting Schrottenloher (2026) and Luo et al. (2026) for the previous state of the art. The run took 53 hours.

For a second problem, Anthropic says Claude redesigned the most expensive step of Decoded Quantum Interferometry, a leading candidate for quantum speedups on optimization problems. It swapped the extended Euclidean algorithm for Berlekamp-Massey decoding, reconstructed temporary values from the circuit's state rather than storing them, and sized its memory from a 403 million sample simulation rather than for the worst case. Anthropic says the resulting circuit uses 12 times fewer gates than Google Quantum AI's previous construction.

The method is the news

Behind both results is one loop. Claude proposes a Python implementation. An evaluator program runs it, checks it for correctness, and scores it on qubit and gate count. The score goes back to Claude, which uses it to improve the next attempt. Anthropic describes the human input as "minimal", meaning the initial brief and occasional prompts to broaden the search.

Anthropic says the setup required Claude to propose several qualitatively different approaches in parallel, and that the researchers prompted it to broaden its search when it was being too conservative. By the research team's own account, the model's default was to prefer well-known approaches and second-guess apparent improvements, which the parallel-approaches rule and the occasional nudge counteracted.

The claim is a resource reduction, in simulation

Each number is Anthropic's own, measured against the best previously published designs cited in each paper. The result lives on paper: it describes how many qubits and gates a future quantum computer would need to run these algorithms, which the Lean proof for divsteps and the published identity derivations let any other team check in the mathematics rather than in a lab.

For a team building any AI-backed product, the eval loop that produced these results is the same loop that scores a RAG answer against a known fact, a code change against a test suite, or an agent trace against a graded rubric. Anthropic's design was ordinary: a brief precise enough to score, a scorer cheap enough to run many times, and a rule that forces the model to try several approaches at once rather than refine one. The method travels further than the subject does.

Source

• Anthropic Research: Designing more efficient quantum circuits with Claude
• Discussion: Hacker News