Anthropic's Claude Mythos 5 Cuts Quantum Circuit Gates by 17x

Anthropic, working with Harvard researchers Susanne Yelin and Mikhail Lukin of QuEra Computing, demonstrated that its Claude Mythos 5 model can autonomously optimize quantum circuits across an eight-task benchmark. Operating in an automated loop where the AI writes Python circuits and receives feedback based on qubit count multiplied by gate count, the model outperformed previously published designs across chemistry, optimization, physics simulation, number theory, and numerical computing.

In its standout modular inversion result—a core component of elliptic-curve point addition—a 53-hour optimization run yielded a design using 17-fold fewer gates with 12% more qubits than prior art. Claude derived new algebraic identities, generated a decoding formula, and verified its logic by formalizing its divsteps proof in Lean. For Decoded Quantum Interferometry, Claude replaced Google Quantum AI's extended Euclidean algorithm decoder with a Berlekamp-Massey structure. By analyzing 403 million sample decodings across 35 rounds, Claude trimmed memory needs from a worst-case 665 qubits down to 19 qubits, achieving a 12-fold gate reduction while maintaining a minor 1 in 10,000 error rate.

For practitioners, these algorithmic reductions directly lower the physical hardware barriers to fault-tolerant quantum computing. Just as historical optimizations to the FeMoco benchmark reduced nitrogenase active site calculations from 50 trillion Toffoli gates to 340 million—a 150,000-fold jump putting runtime within hours on a 3 million physical qubit system—Claude's automated optimizations shorten execution times and lower logical qubit bounds. While human researchers still supplied evaluators and prompted broader search pathways, the experiment highlights how automated feedback loops can streamline quantum software development.