Test models on multiple-choice questions, then compare accuracy & latency