OpenAI’s mysterious Strawberry project reveals its secrets. OpenAI o1 is the latest AI model that excels at reasoning tasks.

In recent days, there have been many rumors about OpenAI’s “secret project” called Strawberry. The company is now launching the new model behind it: OpenAI o1. This model would respond more slowly than, for example, ChatGPT-4o. This is because the model thinks first and goes through a whole reasoning process before presenting an answer. This makes it suitable for solving complex mathematical tasks. OpenAI o1 will be available in preview for ChatGPT Plus users.

“Think before you answer”

The new OpenAI o1 model thinks before it answers. This is how OpenAI introduces its newcomer. The model will therefore react much more slowly than ChatGPT-4o, but OpenAI has its reasons for this. The chatbot checks better whether what it says is actually true.

OpenAI uses the so-called Chain of Thought, which is built into the model. The model excels in reasoning tasks. Since it first writes down a long “chain of thought” before the answer comes, the model can arrive at several answers and therefore make fewer mistakes. This should also lead to fewer hallucinations.

In the image above you can see an example of o1’s reasoning process. When the model is asked a complex mathematical question, it takes time to give an answer. This is also shown: the model needed 54 seconds to do this. In addition, you can see in which step of the reasoning process the model is in this time. The result is a detailed reasoning that indicates how the model arrived at the answer.

Thinking talent

What makes OpenAI o1 stand out is its reasoning tasks. This can certainly be useful for solving more complex problems. The MATH benchmark shows that o1 scores 94.8 percent. That’s a big difference from GPT-4o’s score of 60.3. What is the model not good for? For uploading files or generating images, GPT-4o is still good. The o1 preview is now available, but only for ChatGPT Plus users.