After months of waiting, OpenAI has finally introduced a series of new models called “o1” that excel in advanced reasoning, which was previously called Strawberry AI. The new models include OpenAI o1, OpenAI o1-preview, and OpenAI o1-mini. The Preview and Mini models are available today for paid ChatGPT Plus users. At a later date, OpenAI o1-mini will also be available for free ChatGPT users.
According to OpenAI, the o1 models take some time to think before generating an answer, but they can “reason about complex tasks” and solve more difficult problems in math, science, and coding. Additionally, OpenAI claims that the new reasoning models perform as well as PhD students on difficult scientific topics.
To give you an idea, the OpenAI o1 model scored 83% in a rigorous test like the International Mathematical Olympiad (IMO) while GPT-4o could solve only 13% of the problems. In the Codeforces competition, the new o1 model reached the 89th percentile while GPT-4o was at the 11th percentile.
In the MMLU benchmark, OpenAI o1 scored 92.3 and in the MATH benchmark, it scored 94.8. OpenAI claims that in tasks where advanced reasoning is required, o1 closely matches the performance of human experts, which is quite significant.
The o1 models were trained using a chain-of-thought technique through reinforcement learning. It breaks down steps into simpler ones and approaches each step through different strategies until it reaches the correct conclusion. By the way, currently, the o1 models only support text input. You can’t use the model to browse the web or analyze files and images.
