If you think the recent releases of GPT-4o and GPT-4o mini have kept OpenAI busy enough, then you are mistaken. Just four months after the release of GPT-4o, OpenAI has finally released the rumored "Strawberry" Large Language Models (LLMs), officially named o1 and o1-mini.
On Thursday, OpenAI released the o1 series models, which excel in reasoning and can solve complex mathematical, scientific, and coding problems. This release comes over a year after rumors circulated about OpenAI developing a large language model specialized in reasoning (codenamed "Strawberry Project").
Even OpenAI CEO Sam Altman joined in on the rumors, posting a picture of a strawberry garden on X in August with the cryptic caption "I love summer gardens" as shown below.
I love summer gardens pic.twitterspan.com/Ter5Z5nFMc - Sam Altman (@sama) August 7, 2024
Now, the "Strawberry Project" in the form of o1-preview and o1-mini is finally here. Here's everything we know about these models, how to access them, and what you can expect.
What is the "Strawberry Project"?
In May of this year, Reuters obtained an internal document from OpenAI revealing the development of a project codenamed "Strawberry." In this internal document, "Strawberry" is described as a model with advanced reasoning capabilities that can autonomously and reliably navigate the internet for in-depth research.
This project is similar to the Q model, which had rumors circulating since November 2023. At that time, Reuters reported that some OpenAI employees believed Q* to be a breakthrough in the field of Artificial General Intelligence (AGI).
The official response to these rumors came with the release of the o1-preview and o1-mini models on September 12. Unlike the aforementioned rumors, these models are not AGI or fully autonomous, but they do possess advanced reasoning capabilities that researchers and scholars can utilize. These models are now available to ChatGPT Plus and Team subscribers.
What can o1-preview and o1-mini be used for?
The main selling point of these two models is their proficiency in reasoning, allowing them to think and answer questions like humans. OpenAI states that they train these models to think before answering, improving their thought process and enabling them to try different strategies and identify their own mistakes.
Due to their ability to reason out prompts, these models perform better in solving complex or challenging mathematical, scientific, and programming problems. Therefore, if you need assistance with more complex problems, it is best to choose o1-preview. OpenAI also highlights the effectiveness of o1-mini in programming.
How does o1 compare to GPT-4o?
The advantage of the o1 model lies in its outstanding reasoning capabilities. This is achieved through its training method, which requires the model to think before answering, allowing it to try different strategies and ultimately arrive at an answer.
As a result, the o1 model excels in mathematics, science, and programming. For example, it achieved a score of 83% in the International Mathematical Olympiad (IMO) qualification exam. In contrast, GPT-4o only correctly solved 13% of the problems. Therefore, when it comes to solving complex or challenging STEM (Science, Technology, Engineering, and Mathematics) problems, the o1 model is the best choice.
However, despite this performance, it may lead you to believe that the o1 model is always the better choice. But that is not the case. As an early-stage model, o1 lacks some key features of ChatGPT, such as browsing the internet and accepting media uploads. Therefore, for most requests, GPT-4o remains the better choice.
ChatGPT Plus and Teams users can switch between different models based on prompts, so they are not limited to using just one. To make it easier for users to choose, OpenAI is developing a feature that allows ChatGPT to automatically select the best model based on user prompts.
Where can Project Strawberry/o1-preview and o1-mini be accessed?
ChatGPT Plus and Teams users can access the o1-preview and o1-mini models through the model selector toggle button in the top left corner of the ChatGPT page. For o1-preview, there is a weekly limit of 30 messages, while o1-mini has a weekly limit of 50 messages.
Altman confirmed that this update is available to all ChatGPT Plus/Teams users. Eventually, OpenAI also plans to bring these features to all ChatGPT free users, but they have not specified when this change will occur.
Developers also have the opportunity to access these models, especially those who qualify for API usage tier 5, with a limit of 20 requests per minute (RPM) in the API. ChatGPT Enterprise and Edu users will gain access early next week.
Are o1-preview and o1-mini free?
These models are not currently free, but OpenAI plans to make them accessible to everyone. Until then, the only way to access o1-preview and o1-mini is through paid subscriptions, including ChatGPT Plus, Teams, Enterprise, and Edu. Developers can also access them through the API.
For individual users, ChatGPT Plus is the best way to access these models, with a monthly cost of $20 and additional benefits such as priority access to the latest models and the ability to create GPTs.
