What's the best OpenAI model for the most human-sounding outputs?

Hi @tea, I agree with what has already been said above. The key is a good, solid prompt. For example, if you’re aiming for something “human-sounding,” I’d suggest first thinking through what “human-sounding” means for your use case, then describing the desired model behavior, tone and response style in the Developer message. When it comes to models, my own opinion is that GPT-4o and GPT-4.1 are still very good at holding tone. GPT-5.1, GPT-5.4 and GPT-5.5 can also work well if the prompt is stable and specific enough.