Anthropic reports that Claude Mythos 5 proposed and verified quantum circuit designs on its own inside a closed loop scored by an evaluator program, producing better designs on 8 quantum circuit design tasks, 7 of which beat the previously published state of the art. The research is a preprint and has not yet been peer reviewed.
1. Elliptic-curve point addition: gates down to 1/17. Claude derived a reversible version of the divsteps algorithm: instead of storing the branch information at every step, it reconstructs the branches from intermediate values still sitting in the registers, with the key proof formalized in Lean. One 53-hour run with no human intervention between rounds delivered the result: gate count down to 1/17 of the previous design, using only 12% more qubits.
2. DQI decoder: gates down to 1/12. For the Reed-Solomon decoder of Decoded Quantum Interferometry (DQI), Claude replaced the extended Euclidean algorithm with Berlekamp-Massey decoding and ran 403 million sample decodings to measure real memory needs: just 19 qubits to store a single value, instead of the worst-case 665. Gate count fell to 1/12 of Google's earlier construction, at the cost of a roughly 1-in-10,000 failure rate, within the task's allowed error budget.
3. The human role. Humans supplied only the initial brief and occasional prompts to think more broadly; the evaluator checked correctness and scored the circuits. Anthropic says the same loop could one day design an entire error-corrected quantum computer.
Source: Read the original. This article is a summarized digest for reference.
