Claude finds a new quantum circuit design in a 53-hour search

Anthropic says a 53-hour search with Claude produced a new design for elliptic-curve point addition, an arithmetic routine used in quantum algorithms. The research, announced on October 16, concerns mathematical and code-level circuit designs—not a speedup demonstrated on a quantum computer.

A circuit search with a checkable target

Researchers gave Claude Mythos 5 plain-language specifications for basic routines used in larger quantum algorithms. Claude wrote and revised Python implementations, while a separate evaluator program checked correctness and returned a resource score for the next attempt.

Qubits hold quantum information, and gates are elementary operations performed on them. The evaluator multiplied the qubit count by the gate count, giving Claude a concrete target for reducing the resources a circuit would need.

The paper reports that seven of eight tasks reduced both qubit counts and counts of costly Toffoli gates compared with the best previously published designs. The eighth offered a new trade-off between those resources, rather than reducing both at once.

Recovering the history instead of storing it

A major difficulty is reversibility. In the circuits studied, intermediate working information cannot simply be deleted. Keeping it consumes qubits; undoing the calculations that created it takes additional gates.

For the point-addition task, Claude explored divsteps, a classical method for modular inversion—the wraparound arithmetic equivalent of division. A straightforward reversible implementation would need to retain a long record of the choices made during the calculation.

Claude developed a way to reconstruct that branch history from the current numerical state instead of storing the full record. According to the researchers, it derived mathematical identities behind the method, tested it through large numerical simulations and formalized the decoding claim in Lean.

Fewer gates can still require more qubits

Reconstructing history also takes work, so Claude developed a family of designs that sometimes store information and sometimes recover it. Compared with the smallest previous point-addition design, one uses 12% more qubits but 17 times fewer gates.

That is a gate-count comparison, not evidence of a quantum computer running 17 times faster. It gives designers another option when balancing limited qubit storage against the operations a future machine must perform.

Human guidance and hardware problems remain

The search was not entirely human-free. Normal evaluation rounds ran without human intervention, but researchers supplied the initial brief, required several different approaches to be explored in parallel and occasionally prompted Claude to broaden its search.

These are findings reported by the research team in its announcement and arXiv preprint, not an independent reproduction. More efficient circuits also leave substantial hardware work: fragile qubits still need noise control, error correction and reliable operation.

The useful research pattern is specific: let a coding agent explore designs whose correctness and resource costs can be checked by another program. The paper provides the details for examining that approach; the results do not establish a commercial timetable or show that scientific research can generally run unattended.