Anthropic Says Claude-Designed Quantum Circuits Beat Published Best Results

Anthropic said on Oct 16, 2026, that its AI model Claude designed quantum circuits more efficient than the best published results for eight classes of quantum computing tasks, including chemistry, optimization, physics simulation, number theory, and numerical computation, with almost no human intervention.

Quantum computers rely on qubits, but current machines typically have only hundreds to thousands of qubits and are highly sensitive to heat, vibration, and electrical noise. The number of qubits and quantum gates is a key cost metric, and reducing both can lower hardware demands, runtime, and accumulated noise.

The team described the tasks to Claude in plain language and built an automated evaluation program that checked correctness and scored circuits by qubit count multiplied by gate count. Claude then iterated on feedback. According to Anthropic, the resulting designs were not minor tweaks to known circuits but introduced new mathematical and structural approaches.

Examples include using molecular symmetry and repeated structure when simulating iron sulfide, and cutting the cost of the most expensive step in a quantum optimization decoder to about one-twelfth. In one notable result involving elliptic curve point addition, Claude explored autonomously for about 53 hours and produced reversible algorithms that run backward without storing the full branch history, offering different trade-offs between qubit count and runtime.

The researchers said such advances could reduce the hardware scale required for practical quantum applications, while noting that circuit design is only one part of building fault-tolerant quantum computers.