OpenAI Launches o3, o4-mini Models Featuring Native Agentic Functions
OpenAI has launched two new o-series reasoning models, that is, o3 and o4-mini. The company explained that these are its first AI reasoning models with in-built agentic capabilities that enable them to combine and leverage tools that can be accessed within ChatGPT, including web searching, file analysis, visual reasoning, and image generation. OpenAI explained that the models are trained to determine when and how to apply these agenic tools to produce rich responses in the right format.
In a blog post, OpenAI explained that the new models use their agentic abilities to tackle complex, multi-step questions more effectively. The company sees this as an initial step toward building a true agentic ChatGPT that can independently perform tasks on a user's behalf.
OpenAI’s new agentic reasoning models: Details
o3 model: OpenAI characterized the new o3 model as its strongest reasoning model yet, performing well across all domains like coding, math, science, and visual understanding. The firm explained that o3 is best suited for tackling multi-layered issues where solutions are not readily apparent. It is also reported to perform well on visual tasks like examining images, charts, and diagrams.
o4-mini model: The o4-mini is a smaller model optimised for speed and cost-efficiency. Despite its size, it reportedly performs excellently in maths, coding, and visual reasoning during internal tests. OpenAI said it outperforms its predecessor, the o3-mini, in areas like data science, while also supporting higher usage limits compared to the new o3.
OpenAI’s new agentic reasoning models: Visual abilities
OpenAI emphasized that these models are the first to incorporate images as part of their line of reasoning. As per the company, "they don't just see an image they think with it".
With this capability, customers can post images of whiteboards, textbook illustrations, or hand-drawn charts, and the models are able to comprehend them even if the image is low resolution, reversed, or blurry. With tool use allowed, the models can even manipulate images in real time, with operations such as rotating, zooming, or converting visuals part of their thought process.
Availability
OpenAI stated that o3, o4-mini, and the increased capacity o4-mini-high model are also currently available as part of the model selector menu in ChatGPT for users subscribed to Plus, Pro, and Team plans. Free plan users can however test the latest o4-mini model by clicking 'Think' in the composer before sending over their question. OpenAI added that it planned to introduce an o3-pro model in the near future as well.
