Recently, OpenAI released their new “reasoning” models called o1-preview and o1-mini. Keep in mind that these models are not reasoning and only mimic it.
After some testing, they don’t feel like new models but more like an implementation of Chain-of-Thought (CoT) combined with prompt chaining. OpenAI even advises against using Chain-of-Thought prompts when using the new models.
What are those esoteric techniques, you ask?
Chain-of-Thought is a prompting technique used to help the model mirror human reasoning by asking it to think about the problem step by step.
Prompt chaining is a technique used to generate a desired output by providing a series of prompts instead of one.
What are the costs?
The o1-preview model costs ~4 times more than GPT-4o and ~100 times more than GPT-4o mini.
The o1-mini model costs ~25% less than GPT-4o and ~20 times more than GPT-4o mini.
The costs are higher as the models are doing more intermediate steps. It’s like they are running multiple prompts before they come up with the answer.
Should you switch to the new models?
Currently, only those who spent more than $1000 on the API can use the new models via API.
Even so, I suggest staying with the cheaper models and refining your prompts instead. It should increase the desired outcome without significantly increasing your prices, as using the new models would.