Anthropic handed Claude a narrow job and stepped back. In a research post published October 16, the lab reported that Claude Mythos 5 designed quantum circuits beating the best published versions across eight building blocks of quantum algorithms.
The tasks came from chemistry, optimization, physics simulation and number theory. Each started as a plain-language specification, paired with an automated evaluator that checked correctness and scored each attempt. Claude wrote its designs as Python, read the feedback and revised, with no human in the loop between rounds.
Why the model could compete: circuit designs are code, correctness can be tested by an ordinary computer, and the scoring metric of qubits multiplied by gates gives a clean objective. Those three properties turn the problem into a tight feedback loop.
The gains matter because quantum hardware is scarce and fragile. Fewer qubits and gates mean a smaller machine could run useful chemistry or materials science work, and sooner than previously assumed.
The work, led by Andi Gu under Anthropic’s STEM Fellows Program with Harvard collaborators, also produced a paper submitted to arXiv on October 15. Anthropic published its own summary of the framework, and the eight-task result rests on Anthropic’s evaluation rather than an outside review.
