On September 12, OpenAI previewed its new model, OpenAI o1, designed to handle complex tasks like writing code, solving math problems, and performing deep reasoning. It’s the first in the long-rumored family of next-generation AIs, dubbed “Strawberry.”
ChatGPT Plus and Team users and developers using OpenAI API Level 5 can now access the full model preview, o1-preview.
These users can also access o1-mini, a smaller and faster version of the o1 model that is particularly effective for coding. Because it is a smaller model, the tech giant claims it is “80% cheaper than o1-preview, making it a powerful and cost-effective model for applications that require reasoning but not extensive knowledge of the world.”
Open AI said ChatGPT Enterprise and Edu users will have access to both models starting next week.
“We are also planning to provide access to o1-mini to all ChatGPT Free users,” the company said in its statement.
Here is o1, a series of our most capable and aligned models yet:
o1 is still flawed, still limited, and still seems more impressive on first use than after spending more time with it. image.twitter.com/Qs1HoSDOz1
—Sam Altman (@sama) September 12, 2024
o1 takes longer to reason about more difficult problems
Rather than extending GPT-4’s language capabilities, OpenAI o1 and o1-mini focus on science, code creation and debugging, and math. A demo video shows the model building a playable game in the style of the Snake games of the 1970s. As OpenAI explained, o1 can be used by:
- Health researchers will record cell sequencing data.
- Physicists will generate complicated mathematical formulas needed for quantum optics.
- Developers in all fields to create and execute multi-step workflows.
OpenAI says o1 placed in the 89th percentile on the Codeforces competitive programming test and scored among the top 500 US students in a qualifier for the US Math Olympiad.
By nature, o1 will take longer to respond than ChatGPT or GPT-4.
o1-preview can generate a maximum of 32k tokens, while o1-mini can generate a maximum of 64k tokens. A token can be a single character or a single word in length, depending on the complexity of the text. Both versions of the new model only support text input, not audio or images.
OpenAI has created a best practices guide for developers to determine if o1 is right for their work.
In the model’s system card, where OpenAI describes the efforts of response teams and other security considerations, o1 received a “medium” security rating in two categories. Independent research group Apollo Research noted that o1 “has the basic capabilities needed to perform simple schemes in context” — that is, “gaming its oversight mechanisms as a means to an end.” On the other hand, deeper reasoning gives the model a better understanding of security policies.
