A study published on October 16, 2026, by Anthropic, in collaboration with scientists at Harvard University and quantum computing company QuEra Computing, shows that the AI model Claude Mythos 5 is capable of automatically optimizing quantum computer circuit designs beyond methods previously published in the academic literature.
Led by researcher Andi Gu of Anthropic’s STEM Fellows program, the project established a complete closed-loop evaluator system. In this system, Claude received natural-language problem descriptions, wrote Python source code to simulate circuit designs, and then sent them to an independent testing suite for scoring based on two rigorous metrics: qubit count and quantum gate count. The model operated continuously for 53 hours with minimal human intervention, refining its solutions through each feedback iteration to solve eight foundational circuit problems spanning quantum chemistry, physics simulations, and number theory.
Breakthroughs in modular inversion and optimization algorithms
The greatest challenge in quantum circuit programming is that all operations must be reversible, meaning intermediate data cannot be erased or overwritten as in conventional computers; instead, it must be “uncomputed” to reclaim qubits, which causes logic gate counts to surge.
Claude overcame this barrier by discovering an entirely new modular inversion algorithm for point addition on elliptic curves. Measured results indicate that Claude’s design reduced the quantum gate count by up to 17 times while increasing the qubit count by a mere 12% compared to the previous best-known designs. In another problem involving the DQI quantum optimization algorithm, the model also compressed the Reed-Solomon decoding circuit structure, cutting computational gate counts by roughly 12 times. These fundamentally mathematical improvements significantly reduce the projected runtime of complex algorithms on future hardware.
The gap between software simulation and physical hardware
However, the research team also emphasized a fundamental limitation: all of the above findings remain at the stage of theoretical circuit design and software simulation validation. These solutions have not yet been tested on any physical quantum hardware systems. Building fault-tolerant quantum machines with millions of physical qubits remains a challenging hardware engineering problem that will take years to achieve.
Even so, the study demonstrates a new role for large language models: far from being mere coding assistants, AI can become an autonomous scientific discovery partner when placed in an environment with objective ground-truth criteria and clearly measurable costs.