OpenAI introduces a new model, o1. It is the first in a series of "logic" models designed to answer more complex questions and work faster than humans. A smaller and cheaper version of the O1, the o1-mini, is also being launched at the same time. The O1 model is more efficient at writing code and solving complex problems than previous models, but it is more expensive and slower.
O1 is currently available to ChatGPT Plus and Team users, while Enterprise and Edu users will have this feature starting next week. O1 is available via API and costs $15 per million input tokens and $60 per exit token. In comparison, GPT-4o costs $5 per million input tokens and $15 per output token.
Jerry Tworek, head of research at OpenAI, said that o1 was developed with a completely new optimization algorithm and special training data. It is noted that this model, unlike previous models, learns to solve problems on its own. Thanks to this new training methodology, OpenAI claims that the model is more accurate, but stressed that the "hallucination" problem is not completely solved.
The O1 model is better able to solve and explain complex problems, especially in coding and mathematics. For example, O1 scored 83% on the qualifying exam for the International Mathematical Olympiad, while GPT-4o scored only 13%.
However, O1 is not as strong as GPT-4 in some areas. Has poor results in information literacy and is still unable to access the Internet or process files or images.
