Claude Mythos 5 cut gates and qubits on seven of eight circuit tasks
Anthropic said Friday that Claude Mythos 5, in a closed loop with an automated checker, found quantum circuits that use fewer Toffoli gates and fewer qubits than the baseline on seven of eight tasks. The preprint is arXiv:2610.22367, submitted 15 October, not peer-reviewed. The elliptic-curve design came from a 53-hour run. The “17 times fewer gates” line is one comparison, against the design that used the fewest qubits, and that circuit uses about 12 percent more qubits.
Deep dive4 minIn-depth
Priya Shah4 min read
- Anthropic
- Claude Mythos 5
- Quantum
- arXiv

Photo: AnthropicSource page
Anthropic said on Friday that Claude Mythos 5 had designed quantum circuits that improve on published baselines for a set of small, expensive subroutines. The preprint, “Discovering state-of-the-art quantum circuits with large language models,” was submitted to arXiv on 15 October as 2610.22367. It has not been peer-reviewed. The authors are Andi Gu, of Harvard and Anthropic’s STEM Fellows Program, Alek H. Kemeny of Anthropic, and Susanne F. Yelin and Mikhail D. Lukin of Harvard. Lukin is also with QuEra Computing. Anthropic’s post says Gu led the work in the fellows program, with mentorship from Kemeny, a quantum physicist at the company.
The loop is in the paper. A hand-written specification says what the circuit must compute. The model proposes several approaches and writes Python that builds each circuit. An independent evaluator checks the candidate and scores it. The model sees only the feedback the checker returns. The search runs in OpenEvolve, an open-source evolutionary coding tool the paper cites, and it keeps designs that trade gates against qubits so a circuit that is worse on one of the two can stay in the pool. The score is T times Q: the number of Toffoli gates multiplied by the peak number of qubits in use. A Toffoli is a stand-in for runtime because, on an error-corrected machine, that gate is assembled from a specially prepared resource called a magic state. Google Quantum AI reported on Thursday, in Science, that it had cultivated those states on its Willow processor. The counts in the preprint are resource estimates. The authors check them on an ordinary computer.
Seven tasks, and a baseline that was built for the paper
The abstract says that in seven of the eight tasks the discovered circuits reduce both the Toffoli count and the qubit count relative to the state of the art. Table 1 gives the combined-score ratios where the authors compute one. A fixed-point Gaussian goes from 24,000 Toffolis and 362 qubits to 1,583 and 139, a ratio of 39. A Reed–Solomon decoder used in decoded quantum interferometry goes from 5.1 million Toffolis and 1,938 qubits to 440,000 and 1,837, a ratio of 12. Anthropic’s post describes that decoder result as about twelvefold fewer gates than Google’s earlier construction, from a switch to Berlekamp–Massey decoding, from reconstructing temporary values instead of storing them, and from sizing memory for typical inputs rather than the worst case. An iron–sulfur alias-select step, the smallest numbered gain in the table, goes from 55,000 Toffolis and 139 qubits to 15,000 and 126, a ratio of 4.1.
One of the seven is marked. For the Hermitian decoder the table’s note says there was no published reversible circuit, so the baseline is a straightforward one the authors built for the comparison. SymbiQ, which read the preprint including the supplement, makes the same point. The eighth task is the one the abstract separates: a new reversible way to do modular inversion over a prime field, inside elliptic-curve point addition. The paper says it reports no single baseline and no single gain for that task, because earlier designs trade gates and qubits differently.
Where the 17 times comes from
Anthropic’s post says the elliptic-curve design came from one 53-hour run, with human input limited to the opening brief and occasional prompts to search more broadly. The post’s comparison is specific. “Compared with the smallest previous design, one circuit in the family uses 12% more qubits but 17x fewer gates.” Table S3 is the list behind that sentence. Luo and co-authors use 70 million Toffolis and 838 qubits. The divstep circuit the search found uses 4.0 million Toffolis and 935 qubits. Seventy divided by four is 17.5, and 935 is about 12 percent more than 838. That is the fewest-qubit baseline, not the fewest-gate one. Schrottenloher’s design in the same table uses 1.9 million Toffolis and 1,446 qubits. Against it, the divstep circuit uses more gates and fewer qubits. On the paper’s own T-times-Q score, Schrottenloher’s 2.7 billion is lower than the divstep circuit’s 3.7 billion. SymbiQ’s reading of the table is the same arithmetic.
The paper is plain about what the score leaves out. Qubit count, depth, routing and measurements matter in different proportions once a real error-correcting machine is specified, and the authors say choosing that mix is a judgment about the hardware, not a number the search can settle. Anthropic’s post says the early runs were cautious, that the setup had to require several different approaches in parallel, and that people sometimes told the model to look more broadly. The 53 hours are one task. The checker, the score and the eight specifications were written by people.


