Anthropic says its Claude Mythos 5 model has autonomously redesigned quantum circuits for eight scientific tasks and beaten the best previously published design on every one of them. In the arXiv preprint posted alongside the company’s October 16 announcement, seven of the eight designs cut both Toffoli-gate and qubit counts; the eighth is a new modular-inversion architecture that trades 12% more qubits for 17 times fewer gates.
Nothing was run on a quantum computer. The circuits are resource estimates, checked by an automated evaluator and by simulation, for machines that do not yet exist at the required scale. The paper has not been peer reviewed, and no independent replication is public yet.
How Claude Worked: A Closed Loop With An Automated Judge
The work came through Anthropic’s STEM Fellows Program and was led by Andi Gu with Alek Kemeny, alongside Harvard’s Susanne Yelin and Mikhail Lukin, who is also a co-founder of QuEra Computing. Researchers handed Claude each task as a plain-language specification. The model wrote circuit implementations in Python; a separate evaluator program checked each design for correctness and scored it by qubits multiplied by gates, a proxy for runtime; Claude then revised and resubmitted, with no human input between rounds.
The eight tasks spanned chemistry, optimisation, physics simulation, number theory and numerical computing. Anthropic notes the model was conservative early on — repeatedly re-checking its own results and preferring known approaches — until the setup forced it to explore several qualitatively different strategies in parallel.
The Headline Results: A 53-Hour Run And A 12x Decoder
The longest single effort was elliptic-curve point addition, a benchmark that sits at the heart of quantum attacks on discrete-logarithm cryptography. Claude worked for 53 hours with minimal human input and produced what Anthropic calls the first reversible implementation of divsteps — the classical modular-inversion method — with no stored branch history. Instead of recording every branch it took, the circuit reconstructs that history from the numbers themselves, using new algebraic identities Claude derived and then formally verified in the Lean proof checker. One design in the resulting family uses 12% more qubits than the smallest previous circuit but 17 times fewer gates.
For a decoder used in Decoded Quantum Interferometry, Claude replaced the extended Euclidean algorithm with Berlekamp–Massey decoding and reconstructed temporary information instead of storing it. The resulting circuit used roughly 440,000 Toffoli gates, down from about 5.1 million — a cut of around 12x against the previous construction, which came from Google — while peak qubit use fell from 1,938 to 1,837. On another optimisation task, Claude ran 403 million sample decodings, found that almost every run needed just one saved value, and reserved 19 qubits instead of a worst-case 665, making that step about 12 times cheaper at an error rate of roughly one failure in 10,000 runs, within the task’s allowed tolerance.
Why It Matters
The central entity is Claude Mythos 5’s circuit-design loop, and its key attribute is verifiability: circuit design is expert work where a classical computer can check the answer, which is exactly the terrain where a model in a feedback loop can grind past human designs. Its neighbours include the quantum hardware builders, whose resource estimates for useful machines just got cheaper on paper, and cryptographers, because the elliptic-curve result feeds directly into estimates of what a future quantum attack on today’s encryption would cost — the same week Claude was in the news for tracing the NIST curve seeds to a 1990s clock and lrand48. It also extends the pattern readers will recognise from Claude Code: Anthropic’s models doing long, self-directed technical work with a checker in the loop. The honest caveat belongs in the same paragraph: these are paper estimates, and the gap between a cheaper circuit diagram and a working fault-tolerant machine remains measured in years.
Background
Quantum circuit optimisation has been a slow relay of human effort: the FeMoco molecule simulation that anchors much of the field was estimated in 2017 at roughly 50 trillion Toffoli gates, and eight years of work by the world’s top teams compressed that to about 340 million. Anthropic, founded in 2021 as an AI safety lab, has been pushing Claude into scientific work through its STEM Fellows Program, and this preprint — “Discovering state-of-the-art quantum circuits with large language models” — is its most concrete claim yet that a model can do the compressing itself.
Source: The AI Pin and AI Weekly, reporting Anthropic’s research post and arXiv preprint 2610.22367.
0 Comments