The single-inference visual reasoning latency of mainstream multimodal models such as GPT-4o Vision and Gemini 1.5 Flash is typically between 500ms and 3 seconds, and the full perception-decision-execution loop may exceed 5 seconds.