Gpt3.5 is too dumb. Try llama 70b if you want minimal costs but still good performance