Anthropic says its Claude Mythos 5 model has designed quantum circuits that improve on the best previously published versions, in some cases by an order of magnitude or more. In a research post published on October 16, 2026, the company reported that Claude, working in a closed loop with an automated evaluator and largely on its own, took on eight computational primitives — the basic routines from which larger quantum algorithms are built — spanning chemistry, optimization, quantum simulation, number theory and numerical computing. In seven of the eight tasks, the circuits it found significantly reduced both Toffoli gate counts and qubit counts compared with the previous state of the art.
The setup played to a language model's strengths. Quantum circuits can be written as ordinary code, each design can be checked automatically for correctness, and the score — essentially qubits multiplied by gates — gives a clear, objective target. Claude proposed designs in Python, received correctness feedback and a score from the evaluator, and revised, with no human input between rounds. The work was led by Andi Gu through Anthropic's STEM Fellows Program, with Alek Kemeny at Anthropic and collaborators Susanne Yelin of Harvard University and Mikhail Lukin of Harvard University and QuEra Computing. A companion paper, submitted to arXiv on October 15, runs 46 pages.
The headline result came from elliptic-curve point addition, a standard benchmark whose cost is dominated by one step, modular inversion. Given a one-paragraph brief, Claude worked for 53 hours with minimal human input and arrived at a completely new approach: the first reversible implementation of the classical divsteps method that stores no branch history. Along the way it proved several new algebraic identities, checked them in large-scale simulations and formalized the claim in the Lean proof assistant. One circuit in the resulting family uses 12 percent more qubits than the smallest previous design but 17 times fewer gates.
A second result targeted decoded quantum interferometry, or DQI, described as one of the strongest candidates for a large quantum speedup. Claude replaced the previous decoder's extended Euclidean algorithm with Berlekamp-Massey decoding, then ran 403 million sample decodings to show that memory for exceptional cases could be cut from a worst-case 665 qubits to just 19, at a failure rate of about one run in 10,000 — within the task's allowed error. Together, the changes reduced the decoder's gate count by a factor of 12 relative to Google's previous construction.
Gate counts matter because today's quantum computers have only hundreds to thousands of noisy qubits, and every operation is a chance for error. Leaner circuits bring useful algorithms closer to what real hardware might one day run. Anthropic was candid about the limits: Claude was conservative in early experiments, repeatedly rechecking results and favoring known approaches, and the team had to require several different approaches in parallel and occasionally prompt it to search more broadly. The results are also company-reported designs, not demonstrations on quantum hardware, and independent replication will matter.
Even with those caveats, the pattern is notable. Progress on problems like simulating the FeMoco molecule took many research groups eight years to cut estimated costs roughly 150,000-fold. A single 53-hour automated run will not replace that kind of work, but it suggests AI systems are becoming genuine tools for exploring design spaces that humans can only sample — and that the next gains in quantum computing may come as much from better circuits as from better machines.