OpenAI o1-preview

OpenAI released a new AI model called o1, which the company says is the first in a “new series of reasoning models for solving hard problems.” Is this the rumored project Strawberry? It has many of Strawberry’s rumored features. Let’s dig in.

I’ve been using o1 for a few hours, and it does this quirky thing when you enter a prompt: it tells you it’s “thinking,” then walks you through the steps it’s taking to solve the problem. The model is very slow (compared to GPT-4o), so it needs to do something to pass the time, but “thinking”? It’s not thinking – at least not the way humans do.

This descriptive overreach may come back to haunt OpenAI. When legislators are given a label for something, it usually sticks. I’m not sure the AI industry needs a bunch of laypeople and lawmakers spreading FUD that now there are AI models that think like humans. First, it’s not true. Second, it’s going to become a marketing nightmare. Moving on.

OpenAI admits that they suck at naming their products, so “OpenAI o1-preview” (which I assume will ultimately just be named “OpenAI o1”) should not be confused with GPT-4o. The models are different, and (at the moment) they do different things.

This new release includes o1-preview and a smaller, more affordable version called o1-mini. OpenAI designed o1 to improve on tasks such as writing code and solving multistep problems, surpassing GPT-4o in specific benchmarks like math and coding, although it is slower and more expensive to use. ChatGPT Plus and Team users already have access to o1-preview, with Enterprise and Edu users gaining access next week. The price for using o1-preview via the API is significantly higher than GPT-4o: $15 per million input tokens and $60 per million output tokens.

According to OpenAI, o1 was trained using reinforcement learning – a technique that rewards correct behavior and penalizes incorrect responses – and it processes queries step-by-step, “similar to human reasoning.” OpenAI claims the model hallucinates less than previous versions, but it still hallucinates a bit. Notably, o1 scored 83% on an International Mathematics Olympiad exam; GPT-4o scored 13%.

The company notes: “This is an early preview of these reasoning models in ChatGPT and the API. In addition to model updates, we expect to add browsing, file and image uploading, and other features to make them more useful to everyone. We also plan to continue developing and releasing models in our GPT series, in addition to the new OpenAI o1 series.”

I’m only a few hours into using o1. So far, it seems like o1 has less knowledge than GPT-4o, but is better at problem solving; this is what OpenAI says about it, too.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.