One input and one output without any back and forth. Not using for continuous conversation

Hello, I am using ChatGPT API to feed a prompt (70ish words) and use whatever response it gives back (400ish words). No conversation. I also don’t really care for latest broader knowledge. It is a very simple use: one input and one output without any back and forth. What model should I use? gpt-3.5-turbo-0125 or gpt-3.5-turbo-instruct?
Thank you so much for your help.

All GPT-3.5-turbo models have the same knowledge cutoff date.

No models have a built-in chat history; their default operation is stateless independent inputs, and you’d have to provide anything you want the AI to know about previous interactions yourself with each API input. However, you might be talking about the nature of the chat models that want to add the voice of their own narrator unless you write a system message with considerable instruction of performing a role other than being a chatbot.

gpt-3.5-turbo-instruct is a completion model that runs on the completion endpoint. It has a bare prompt, and less chat-like training. That also means that some of the writing skills are different and you will get different length and quality of responses.

Here’s an API playground preset to show an example of the different prompting style used for the instruct model (which has enough training to terminate its output instead of writing forever). Creative story writer. Note the temperature and top_p settings are adjusted so it doesn’t write the same thing twice.

Yeah, I don’t care for chat history or previous interactions. Each input and its gpt output are treated completely independently from other inputs/outputs. So there shouldn’t be any difference in instruct vs 0125? If it’s a HUGE difference in speed for returning output of about 400ish words, that matters. Otherwise, I can just go with 0125 which is cheaper than instruct.

gpt-3.5-turbo-0125 is a chat model. Like ChatGPT.

gpt-3.5-turbo-instruct-0914 is a completion model. Like autocomplete on your phone.

If your output is a continuation of the input, gpt-3.5-turbo-instruct-0914 might be a better choice. If you’re gonna use it like chatgpt, gpt-3.5-turbo-0125 is gonna be a better choice.

One’s a bic razor, the other’s a razor blade. Which one should you use?

It depends on what you’re trying to do! But it should be fairly obvious if you try em on the playground, like _j linked.

Thank you to both of you. That made my understanding of models much better. And since I knew what to ask, I asked ChatGPT with added question “since gpt-3.5-turbo-instruct is 3 times more expensive than gpt-3.5-turbo-0125, can I use gpt-3.5-turbo-0125 for my scenario above”, ChatGPT answered it well suggesting that in my scenario above, both models would be fine since my requirements are simple and so I could go with cheaper gpt-3.5-turbo-0125.

I also played in playground and both models were successfully returning a good response as per my needs.