It can be faster and more effective to fallback to a smaller model (gpt3.5 or haiku), the weakness of the prompt will be more obvious on a smaller model and your iteration time will be faster