OpenAI has made o1-mini, the smallest version of its new reasoning family of artificial intelligence models, available for free in ChatGPT. You can even use it without signing up for an account.
Unlike other AI models, such as GPT-4 and Anthropic's Claude, o1 can take a prompt and work through a problem step-by-step before responding with its final output. This significantly reduces error rates or "hallucinations" and allows for much longer, well-thought-out responses.
There are some limitations to using the o1 family of models. It takes significantly longer to get a response, and there is no multi-modal capability, so you can't give it a document or an image. It only works with text.
Giving it the name "mini" is also somewhat misleading. While it is a much smaller model than the o1-preview, in some respects, its performance is better as it was trained on a more modern architecture.
As with the bigger GPT-4o, while o1-mini is available to free users, it does have rate limits, and you will find that you run out of them very quickly. In fact, these rate limits also apply to paid subscribers; those paying get a few more per week.
What's new in o1-mini and preview?
We appreciate your excitement for OpenAI o1 and we want you to be able to use it more.For Plus and Team users, we have increased rate limits for o1-mini by 7x, from 50 messages per week to 50 messages per day.o1-preview is more expensive to serve, so we’ve increased the rate…September 17, 2024
When OpenAI first launched the o1 family of models last week, they said that users would only get 30 weekly messages to test it out. This is because it is a significantly more expensive model to run, as it goes over multiple prompts repeatedly until it gets to the response that it feels is the right one.
Unlike other models, it is also able to generate significantly larger output, including entire codebases and multi-page reports, as we discovered during our early testing.
OpenAI imposed limits on how often you can use it. Because of demand and people still trying to work out exactly what it's useful for, they have since significantly increased those limits over the initial 30 per week.
First, OpenAI completely refreshed everyone's limits, so as of Friday, everyone has a completely clean slate. Then, the company announced it would be increasing the number of messages per week from 30 to 50 for o1-preview and to more than 200 for o1-mini (if you're a paid subscriber).
How to prompt o1-mini
If you are approaching one of the o1 family of models for the first time, you need to take a slightly different approach than you would with GPT-4o or Gemini.
With those models, you tend almost to treat it like a conversation with Google: throw a question at it, add some extra details and hope for the best. However, you need to be more specific because the o1 family of models reasons through their responses.
In one test, I was fairly vague and tasked it with creating a "gamified task manager app in iOS." Asking that question to ChatGPT would have resulted in all of the code to run the app, at least at a very basic level. However, o1-mini gave me a step-by-step design document for producing the app and an outline of the structure for the codebase.
So, if you decide to use it to try and create an app, you must ensure that you’re giving it specific and explicit instructions to provide the code necessary to develop the app. And that level of specificity translates across all use cases.
