Claude Mythos 5 beat the best published human designs on seven of eight quantum circuit benchmark tasks, reducing both qubit count and Toffoli gate count on each, and produced a new reversible architecture with better overall tradeoffs on the eighth.
The largest improvement was a 17-fold gate count reduction on an elliptic-curve circuit, at 12% more qubits than the smallest prior design. Harvard physicist Andi Gu ran the evaluation through Anthropic's STEM Fellows Program; Anthropic frames the findings as task-specific results suggesting AI can aid circuit search, not as evidence that practical quantum computing is near.