Claude Mythos 5 Designs Leaner Quantum Circuits

Anthropic says its most capable model can now design quantum circuits better than the best ones people have published. In a research post dated 16 October, the company describes Claude Mythos 5 working through plain-language specifications for eight circuits used in chemistry, optimisation, physics simulation and number theory, and returning designs that need fewer qubits and fewer gates than the previous state of the art, in some cases by an order of magnitude. The work was led by Andi Gu through Anthropic’s STEM Fellows Program, with Anthropic physicist Alek Kemeny and Harvard’s Susanne Yelin and Mikhail Lukin, and is written up in a 46-page paper on arXiv.

The setup is what made it possible. Claude wrote each candidate circuit as Python code, and a separate evaluator program checked that it computed the right answer and scored it by the number of qubits multiplied by the number of gates. Claude then used that score for its next attempt, with no human input between rounds. Checking is cheap here because these circuits compute ordinary classical functions, so a normal computer can confirm every output. People still wrote the specifications and the evaluator, and they had to push: in early runs, Anthropic writes, Claude was “conservative by nature”, second-guessing real improvements and drifting back to familiar approaches, so the team made it explore several approaches in parallel and occasionally prompted it to think more broadly.

A decoder 12 times cheaper

The clearest result is in decoded quantum interferometry, or DQI, one of the strongest candidates for a large quantum speed-up on an optimisation problem. Estimates from Google Quantum AI imply that one instance of that problem would take about three months on the Frontier supercomputer and roughly an hour on a future quantum computer, with most of that hour spent in a single step, a Reed–Solomon decoder. Claude swapped the decoding method for one that runs a fixed number of identical rounds, then wrote a fast simulator and ran 403 million sample decodings to find out how much working memory it really needed. Reserving room for the worst case would have taken 665 qubits; Claude reserved 19 and accepted that about one run in 10,000 fails, within the allowed error budget. The finished decoder uses about 12 times fewer gates than Google’s construction.

A 53-hour run on elliptic curves

The eighth task was point addition on a 256-bit elliptic curve, the operation at the heart of quantum attacks on elliptic-curve cryptography; the parameters in the paper match secp256k1, the curve behind Bitcoin’s signatures, although the paper does not name it. From a one-paragraph brief, Claude worked for 53 hours with little human input and adapted a classical inversion method called divsteps so that the circuit no longer has to store a record of every branch it takes, recovering it from the numbers themselves instead. It proved the identities that make this work, tested them in large simulations and formalised the key claim in the Lean proof assistant. The paper calls it, to the authors’ knowledge, the first reversible divstep inversion circuit over a prime field.

That result needs reading carefully. Anthropic’s post says one circuit in the new family uses 12% more qubits but 17 times fewer gates than the smallest previous design, which matches the paper’s comparison with Luo et al.: 935 qubits and 4 million Toffoli gates against 838 qubits and 70 million. Against Schrottenloher’s design it needs far fewer qubits, 935 against 1,446, but about twice as many gates. So the post says Claude beat the prior best on both counts across all eight tasks, while the paper’s own abstract says that happened in seven, and describes the eighth as a better trade-off between gates and qubits rather than a win on both.

What it does not show

Anthropic frames the work against a longer history. Estimates for simulating FeMoco, the iron-molybdenum cluster that bacteria use to fix nitrogen, fell from about 50 trillion Toffoli gates in 2017 to about 340 million, a drop of roughly 150,000 times that took many research groups eight years. The company’s argument is that loops like this one could speed that kind of progress up, and that useful quantum computing could arrive sooner as a result. The paper lists the limits: the score ignores fault-tolerance overheads such as routing and magic-state production, several circuits accept a small chance of failure, and some savings only apply to one specific problem. None of the designs has been reproduced by another group yet. In September Claude turned up a CRISPR-like enzyme system in phage DNA; this time the field it touches is the one that decides how soon quantum computers become a threat to the cryptography in use now.