ChatGPT o3-mini

Openai has released for free for all O3-Mini, the new AI model: how it works

After anticipating this last December, Openai announced the Release of O3-Minia new artificial intelligence model designed for advanced reasoning. This model, according to Sam Altman’s company, would represent a remarkable step forward compared to the classic linguistic models, as it has been developed to manage complex problems that require in -depth analyzes. Despite being a more compact and efficient version of O3, which is the top model, O3-Mini maintains advanced capacity in mathematics, science and programming, being particularly useful in the field Stem (Science, Technology, Engineering, Mathematics). Among its distinctive characteristics we mention the use of simulated reasoninga technique that allows the model to “reflect” on your answers before generating them, improving the accuracy. In addition, it introduces an innovative security system called deliberative alignmentwhich allows you to evaluate users’ requests in a more sophisticated way, reducing both the risk of inadequate responses and the unjustified refusal of valid content. Interesting fact, Openii made O3-Mini accessible to users Free, in a period when he tries to keep up with the Chinese rival Deepseek.

The features of 03-mini: how it works

The most important feature of O3-Mini is that This model focuses on reasoninga capacity that makes it more effective in tasks that require logic and deduction. The concept of simulated reasoning represents a significant evolution with respect to previous methods, such as the Cot (Chain of Thought), as it allows the model to analyze a problem more deeply before providing an answer. This approach improves the quality and precision of responses, making it particularly useful in academic and technical areas.

The model introduces one more refined safety management thanks todeliberative alignmentwhich uses the reasoning skills to evaluate the context of a request. Instead of applying simple predefined filters, as happens in traditional systems, the model is able to identify ambiguous intentions or attempts to manipulate, improving protection against improper uses without unnecessarily limiting access to legitimate information.

For developers, O3-Mini also introduces other interesting features, such as the Function call and structured outputs, which facilitate their integration into practical applications. It is also possible to select between three levels of reasoning (low, medium and high), allowing you to balance precision and speed according to your specific needs.

To summarize the Main features of O3-Minitherefore, we can say that compared to the previous model, or O1-Mini, is faster and efficient (according to some analyzes would respond 24% faster without sacrificing high performance in the STEM subjects) and is more accurate (Based on some tests it would provide preferred answers by users in 56% of cases).

As Openai specified, moreover, “O3-Mini now works with research to find updated answers with relevant web sources connections“And added that”This is an initial prototype while we work to integrate research into our reasoning models».

Performance side, O3-Mini proves to be superior to its predecessor O1-Mini in various reference tests. For example, it reaches aaccuracy of 96.7% in the anima (American Invitational Mathematics Examination), compared to 83.3% of O1. In the field of programming, he gets the 71.7% precision on Swe-Bench Verifieda coding benchmark based on real problems, improving by about 20% compared to O1. Even in the scientific field, it exceeds the previous model with a score of the87.7% in the Diamond GPQA testwhich evaluates the ability to answer advanced academic questions. Also demonstrates a higher adaptation capacity in Arc-AGI benchmarkwhich measures the ability of a AI to recognize schemes and solve new problems, obtaining a score of the87.5%higher than the human average of 85%.

Image
The graph highlights a low reasoning effort of the O3 -Mini model in the AIME test (2024), reaching comparable performance with O1 -Mini, while with a medium effort, O3 -Mini reaches comparable performance with O1 and, with a high reasoning effort , is even able to overcome both O1 -Mini and O1. Credit: Openai.

Where and when O3-Mini Openi will be available

The new O3-Mini is already available for chatgpt plus, team and pro userswhile integration for users Enterprise is planned by February. Compared to the previous models, Openai provides for a higher limit of messages: Plus and Team users will see their limit from 50 to 150 messages per day tripled; Pro users will have no limit of use. Another interesting news concerns the possibility of Try O3-Mini for free also on the basic chatgpt levelselecting the option Start the reasoning in the messages of messages. As Openii pointed out in the announcement “This is the first time that a reasoning model has been made available to free users in chatgpt».