Claude Mythos 5 Cuts DQI Decoder Gate Count 12x vs Google

Claude Mythos 5 beat the best previously published quantum-circuit designs on eight computational primitives, including a Reed-Solomon decoder for Decoded Quantum Interferometry whose gate count came in 12 times smaller than Google's previous construction, according to Anthropic.

The headline numbers sit inside a specific setup. Claude ran in a closed loop with an evaluator program that scored qubit and gate counts and iterated without human intervention between rounds. In the elliptic-curve point-addition family, "one circuit in the family uses 12% more qubits but 17x fewer gates" than the smallest previous design, the post reports, calling it the result of a single 53-hour run.

On the DQI decoder, Anthropic says Claude swapped the extended Euclidean algorithm for Berlekamp–Massey decoding and then empirically pruned the memory budget. "Claude ran 403 million sample decodings to measure how often these exceptions actually occurred," the write-up reads, which let it set aside room for just one saved value, or 19 qubits, in place of a worst-case 665. Failures land near 1 in 10,000, which the authors call acceptable. The 12x gate reduction is measured relative to a construction credited to Google, which the post links directly.

The work was led by Andi Gu through Anthropic's STEM Fellows Program, with mentorship from Alek Kemeny, a quantum physicist at Anthropic, and in collaboration with Susanne Yelin and Mikhail Lukin of Harvard, the latter also affiliated with QuEra Computing. On the chemistry side, "when compiling a circuit for simulating an iron-sulfide molecule, Claude noticed that the molecule's symmetries and repeating structure meant that some results could be calculated once and reused," the authors report, part of an arc from 2017 estimates for the FeMoco simulation that the post frames as "a roughly 150,000-fold reduction" in gate counts.

Claude's limits show up in the write-up's own hedges. The authors note the model repeatedly second-guessed its results, and they say they had to force parallel exploration of qualitatively different approaches before it would commit to a direction.