there's a lot of room for optimisation via settings for gpt-5 models, and it remains for vapi to either do the optimisation or expose the parameters for us to do so