AI agents benchmarked on quantum circuit visual understanding with cost-accuracy tradeoffs

Researchers introduced Quantum Circuit Vision, a cost-aware evaluation framework testing whether multimodal AI agents can visually comprehend quantum circuit diagrams and generate verified executable code. The benchmark spans 132 circuits across 13 categories (1-10 qubits) with executable Amazon Braket code and unitary-fidelity verification. Evaluating three frontier Claude-family models at different capability-cost tiers over five repeated trials, the mid-tier model (Sonnet 4.6, 1.30× credits) achieved a 91% pass rate on the core subset at 18% of the per-call cost of the strongest model (Opus 4.6), whose accuracy advantage was not statistically significant (paired t-test, p=0.083). Circuit depth, not qubit count, was the primary predictor of failure (p<0.001). Chain-of-thought prompting showed no statistically significant effect (all p>0.18), suggesting visual pattern recognition outweighs explicit reasoning for structurally coupled diagrams. A cascade routing strategy (cheap → expensive models) reached 84% accuracy at 38% of single-model cost, demonstrating that model routing dominates prompt engineering as a cost lever. The QCV-Dataset (132 circuits, five modalities, 1,931 files) and all evaluation code, cost logs, and verification scripts were released on Hugging Face Hub and GitHub.