Claude Mythos 5 designs quantum circuits with 17x fewer gates in 53 hours

Anthropic has published research showing that Claude Mythos 5 helped design more efficient quantum circuits across eight scientific computing tasks. In one particularly challenging case, the model spent 53 hours developing a new reversible algorithm relevant to elliptic-curve arithmetic. One resulting circuit uses 17 times fewer Toffoli gates than a previously published design optimized for minimum qubit count, while requiring approximately 12 percent more qubits.

Eight tasks, with automated verification

The researchers, led by Andi Gu through Anthropic's STEM Fellows Program in collaboration with scientists at Harvard University, gave Claude written specifications for computational primitives used in chemistry, optimization, quantum simulation, numerical computing, and number theory. Claude proposed implementations in Python and an independent evaluator checked their correctness and scored resource requirements. The model then revised its proposals using the feedback.

According to the accompanying research paper, seven of the eight tasks yielded designs that reduced both the number of qubits and the number of Toffoli gates compared with earlier published methods. The remaining task produced a new trade-off between qubit use and gate count rather than an unqualified reduction in every resource.

Why the 53-hour result matters

The longest highlighted run tackled elliptic-curve point addition, an operation used repeatedly in quantum algorithms for the elliptic-curve discrete logarithm problem. A demanding part of that operation is modular inversion. A conventional algorithm called divsteps works efficiently on ordinary computers, but converting it into a reversible quantum computation has been difficult because the circuit cannot simply erase intermediate decisions.

Claude developed a method to reconstruct the history of those decisions from the current numerical state instead of storing it in separate qubits. The researchers say the model derived new algebraic identities, tested candidate formulas through simulations, and formalized a key correctness argument in Lean. Human input during the 53-hour run was limited mainly to the original task description and occasional requests to explore broader approaches.

What the 17-fold figure actually measures

One design in the resulting family of quantum circuits needs about 12 percent more qubits than the most qubit-efficient previously published design, but roughly one-seventeenth as many Toffoli gates. That is a substantial improvement in one circuit's resource trade-off, not evidence that breaking real-world cryptography has become 17 times cheaper overall. Other published designs can use fewer gates at the cost of substantially more qubits.

The research does not show that Claude recovered a cryptographic private key, broke an encryption system, or ran an attack on a working quantum computer. Fault-tolerant quantum hardware, error correction, and the complete algorithm would still determine the feasibility of any future attack.

A second result: a more efficient quantum decoder

The team also reports that Claude redesigned a Reed-Solomon decoding routine used in decoded quantum interferometry, a proposed quantum optimization technique. By changing the algorithmic approach and avoiding expensive worst-case storage, the design cut the decoder's Toffoli gate requirements by about a factor of 12 compared with a previous construction. To guide that change, the model ran 403 million sample decodings. Its resulting design permits a small failure rate consistent with the task's assumptions.

What has and has not been demonstrated

These are algorithm and circuit-design results evaluated through software checks, numerical experiments, and mathematical verification, not performance measurements on a large fault-tolerant quantum computer. The findings are reported by Anthropic and its collaborators in a research preprint, and further assessment will depend on independent scrutiny and future hardware implementations.

Why it matters

The broader significance is that a language model can participate in discovering new mathematical approaches to a difficult scientific engineering problem, rather than only drafting code from a known solution. If these methods generalize, AI-assisted design could help reduce the hardware requirements for future useful quantum algorithms.

Sources: Anthropic research announcement and the accompanying research preprint .

Source: Anthropic