Claude spent 53 hours alone on a quantum math problem and came back with circuits that beat the published best

Anthropic on Friday, Oct. 16, 2026, reported that Claude, working largely on its own against an automated scorer, found quantum circuit designs that use fewer qubits and fewer gates than the best previously published ones across eight standard tasks.

Quantum hardware gets the headlines, but most of the progress toward useful machines has come from people redesigning the math to need fewer gates. Now an AI did a chunk of that redesign itself, over two days, and checked its own proof in Lean. The same day Anthropic showed Claude untangling where NIST's curve seeds came from, it was also making the circuits that would one day attack elliptic curves cheaper. Nobody's breaking crypto this year, but post-quantum migration timelines just got another reason not to slip.

Anthropic published "Designing more efficient quantum circuits with Claude" on Oct. 16, 2026, with a paper on arXiv by Andi Gu, Alek H. Kemeny, Susanne F. Yelin and Mikhail D. Lukin. The work was led by Andi Gu in Anthropic's STEM Fellows Program, mentored by Anthropic quantum physicist Alek Kemeny, with Yelin at Harvard and Lukin at Harvard and QuEra Computing. Those lines are Anthropic's.

Researchers gave Claude plain-language specs for eight circuits drawn from chemistry, optimization, physics simulation, number theory and numerical computing, plus an automated evaluator that checks each design and scores it on qubits multiplied by gates. A qubit is the basic unit of quantum information. A gate is one of the simple operations a quantum computer applies. Those lines are Anthropic's.

Claude Mythos 5 worked in a closed loop — propose a design, get a score, improve — with no human input between rounds. Across all eight primitives, Anthropic says it found designs using fewer qubits and fewer gates than the best previously published ones, by an order of magnitude or more in some cases. An order of magnitude means about ten times. Those lines are Anthropic's.

On elliptic-curve point addition, Claude worked for 53 hours from a one-paragraph brief, with occasional prompts to think more broadly. It produced the first reversible version of the "divsteps" modular-inversion method that does not store its branch history, by deriving a rule to recover that history from the numbers themselves. It proved new algebraic identities, checked the rule with large simulations and formalized it in Lean, a language used to check mathematical proofs. Those lines are Anthropic's.

One resulting point-addition design uses 12% more qubits but 17x fewer gates than the smallest previous design. The family of designs offers qubit-versus-runtime trade-offs that weren't available before. Those figures are Anthropic's.

For decoded quantum interferometry, or DQI — a leading candidate for quantum speedups on optimization problems — Claude made the algorithm's costliest step, a Reed-Solomon decoder, about 12x cheaper. A decoder, here, is the routine that recovers errors from a compact summary. Those lines are Anthropic's.

Anthropic's context: estimates for simulating the FeMoco molecule, the iron-sulfur cluster at the heart of the enzyme that turns nitrogen into ammonia, fell from about 50 trillion Toffoli gates in 2017 to about 340 million, mostly from better circuit design rather than better hardware. A Toffoli gate is one of the expensive operations used to estimate how long a quantum job will take. Anthropic says improvements like these mean practical quantum computers could arrive sooner than previously anticipated. Those lines are Anthropic's.

Point addition is the core step of quantum algorithms for the elliptic-curve discrete-log problem, the math that protects a large share of today's public-key cryptography. Cheaper circuits lower the bar for a future machine to attack it. No quantum computer today can. This work is not a claim that Claude broke any encryption.

In plain terms, Anthropic said on Friday that Claude Mythos 5, working mostly alone against an automated scorer, found smaller, faster designs for eight standard quantum circuits, including a 53-hour run on elliptic-curve point addition that it then proved in Lean. The company says some of the gains are ten times or more. Nobody is breaking today's crypto with this.