By Hamza Faisal
Anthropic is teaching Claude to think like a quantum engineer — and the results are startling. In research published on October 16, the company says its Claude Mythos 5 model autonomously designed and refined quantum circuits across eight benchmark tasks, beating the best previously published designs on seven of them. In the standout case, Claude produced a circuit for elliptic-curve arithmetic that needs about 17 times fewer gates than the most compact design ever published.
Quantum computers are still noisy, error-prone machines. Every operation — every "gate" — is a chance for an error, so designing circuits that do the same math with fewer gates is one of the field's central engineering problems. Anthropic's experiment asks whether an AI model can do that design work itself.
53 hours, no human in the loop
The setup was a closed generator–judge loop. Researchers gave Claude plain-language descriptions of eight computational tasks spanning quantum chemistry, optimization, physics simulation, number theory, and numerical computing. The model proposed circuit designs in code; an automated evaluator checked each one for correctness and scored it on Toffoli-gate count and peak qubit use; Claude then improved the next version using that feedback — with no human intervention between rounds.
The project was led by Anthropic STEM Fellow Andi Gu, working with collaborators from Harvard University and QuEra Computing, and mentored by Anthropic quantum physicist Alek Kemeny. Designs were ranked by the product of gate count and qubit count — a rough proxy for how hard a circuit would be to run on future hardware — and key claims were validated with numerical simulations and formal proofs in the Lean proof assistant.
The 17x breakthrough — and a 12x win over Google
The headline result came from elliptic-curve point addition, a core step in the discrete-logarithm algorithms a sufficiently powerful quantum computer could one day run. Claude developed a new reversible method for modular inversion — the most resource-intensive elliptic-curve operation — in a single 53-hour run started from a one-paragraph brief. One resulting design used 4 million Toffoli gates and 935 qubits, versus 70 million gates and 838 qubits before: 12 percent more qubits, but 17 times fewer gates.
Claude also redesigned the decoder for Decoded Quantum Interferometry, a quantum optimization approach, coming in 12 times leaner on gates than a construction credited to Google. The model swapped the extended Euclidean algorithm for Berlekamp–Massey decoding, then ran 403 million sample decodings to measure rare edge cases — enough data to keep room for just one saved value (19 qubits) instead of a worst-case 665. Anthropic says the resulting failure rate of about 1 in 10,000 sits within the allowed error budget, making the most expensive step of the algorithm a little over an order of magnitude cheaper.
What it means — and what it doesn't
Anthropic is careful about the limits: these are resource counts on paper, not circuits run on real hardware, and better circuits don't remove the need for better error correction. But the experiment suggests AI could accelerate a field long driven by small teams of specialists — one model explored a design space in 53 hours that would take humans far longer, finding trade-offs that expand the engineering menu as hardware improves.
It's also a glimpse of how AI labs now test their own models: Anthropic put Claude inside the design loop, proposing, testing, and revising against a hard physical metric. The result isn't just a smarter chatbot — it's a machine that did real engineering.
By Hamza Faisal
