Is there any variation in reply when your ran in gpt3.5 and gpt4 ?