OpenAi had a lot to say for the end of 2024. For several days, the most famous AI company in the world revealed its work aimed at improving its ChatGPT artificial intelligence. Among them, the new o3 intelligence model promises to push the limits of everything we have known until now. Here are 4 things to know about o3.
o3 can really “think”
Gone are the days of instantaneous and sometimes approximate answers. o3 is distinguished by their ability to “think” before responding. Unlike traditional models that operate like parrots, o3 uses a “private thought chain” to analyze issues from all angles.
Even more interesting, users can now choose between three levels of reflection : low, medium or high. The higher the level of reflection, the longer the model may take to formulate a response. In return, o3 takes the time to check the facts and avoid the pitfalls that usually trip up other AI models.
Performances that defy belief
The numbers are dizzying: o3 shatters records on virtually every standardized test. He scored a stunning 96.7% on the AIME American Mathematics exam and achieved a score of 87.7% on doctoral-level science questions. In programming, he ranks among the top 1% of global developers with a Codeforces rating of 2727, a level that only the best programmers in the world reach.
But the most impressive remains its performance on the ARC-AGI benchmark, where it reached 87.5% in high performance mode. Designed to assess an AI system’s ability to learn new skills outside of its training data, this test puts o3 at a level never before achieved. However, this exceptional performance comes at a significant cost: each query in high performance mode can cost several thousand dollars in computing power.
o3-mini, the more accessible little brother
OpenAI also revealed o3-mini, a lighter version of the main modelwhich offers remarkable flexibility for everyday tasks. Although more modest than its big brother, it already surpasses the o1 model in many aspects. Its particularity? It allows you to adjust the thinking effort according to the complexity of the task, thus offering an optimal balance between speed and precision.
o3 and o3 mini are not yet available
The release of o3-mini is scheduled for the end of January 2025followed shortly after by that of o3. This gradual deployment strategy demonstrates the caution of OpenAI, which wants to ensure that these new models respect its security principles thanks to a new technique called “deliberative alignment”.
Because while these models are more powerful, they also raise important questions about their potential ability to deceive users, an aspect that the company takes very seriously.
- o3 introduces a three-speed thinking system, allowing “thinking” time to be adapted as needed
- Model performance exceeds expectations, even surpassing human capabilities in some areas
- A more accessible mini version offers an excellent compromise between power and efficiency, available from January 2025
