Every opening phrase is a platitude like ‘sure let’s do it’. So the OpenAI latency is probably higher, they are just using clever orchestration to generate some filler tokens to make latency lower. Unlikely the initial response at OpenAI is coming from the main model.