OpenAI launches o1 model: more time to respond, greater accuracy | OpenAI Playground | OpenAI Login | ChatGPT login | Turtles AI

OpenAI has unveiled its new o1 model, a family of AI models designed to improve reasoning and fact-checking. Despite some initial limitations and high costs, o1 represents a significant advancement over previous models such as GPT-4o. The model is available for ChatGPT Plus and Team users, with expansion planned to Enterprise and Education users.

Key Points:

• Enhanced Reasoning: o1 stands out for its ability to think more deeply and methodically before responding.
• Model Family: The o1 line includes versions such as o1-preview and o1-mini, with limited access to messages.
• Better Technical Performance: o1 excels at complex programming, math, and science tasks.
• High Cost: The model is significantly more expensive than GPT-4o, making it less accessible for general use.

OpenAI recently launched a new line of AI models called o1, aimed at improving the reasoning capabilities of its generative systems. This development marks a significant evolution from the GPT-4o model, with a particular focus on complex tasks such as mathematical problem solving, code analysis, and scientific tasks. The model, previously codenamed “Strawberry,” aims to overcome previous limitations by introducing the ability to “think” for a longer period of time before providing answers, thus optimizing the quality of those answers.

The o1 family currently includes two versions: the more robust “o1-preview” and the “o1-mini,” designed for lighter tasks and focused on code generation. Users who want to try these models must be subscribed to the ChatGPT Plus or Team plans, with an expansion planned for the enterprise and education plans in the future. Free availability for o1-mini is planned, but no specific date has been set.

A key element that distinguishes o1 from its predecessors is its use of an internal reasoning chain, or the ability to perform several verification steps before reaching a conclusion, thus reducing errors. This type of approach is particularly useful for tasks that require complex sequential processing, such as synthesizing information from multiple sources or solving structured problems. The reinforcement learning approach adopted for o1 allows the system to improve through penalties and rewards, making it more accurate over time.

o1’s performance is evident in various tests: in the "International Mathematics Olympiad", o1 successfully solved 83% of the problems, a huge jump from the 13% achieved by the GPT-4o model. Additionally, the model achieved a high percentage in online programming competitions, demonstrating significant improvements in coding and logical reasoning skills. However, these advances come at a cost: using the o1 API is three times more expensive for input and four times more expensive for output than GPT-4o, making o1 less accessible to budget-conscious users. This, along with the weekly message limits (30 for o1-preview and 50 for o1-mini), could be a barrier to widespread adoption in the short term.

Despite its improvements, o1 is not without its flaws. Some tests have shown that the model can still make mistakes or “hallucinate” answers, inventing details that aren’t in the data it provides. These errors are more common than with GPT-4o, but OpenAI has said it plans to continue refining the model over time, with future versions potentially “thinking” for longer periods of time, up to days or weeks, to further improve its reasoning abilities.

As competition in the AI ​​industry continues to heat up, OpenAI faces the challenge of continuing to innovate, improve accessibility, and reduce the cost of its models, while other companies, like Google DeepMind, are taking similar approaches to improving their AI’s reasoning abilities.

With the debut of o1, OpenAI marks a major step toward more advanced AI models that could transform several industries, but only time and its capabilities will tell how effective the model will be.