OpenAI today unveiled two advanced models, "o3" and "o4-mini," part of its series of reasoning models that enable AI to perform logical thinking before providing answers.
This news follows a series of updates recently announced by CEO Sam Altman, in which we anticipated the launch of these two versions.
Capabilities and Features of o3 and o4-mini.
OpenAI states that the "o3" model represents the pinnacle of the company's achievements in reasoning, programming, mathematics, science, and visual understanding. The company described it as its most advanced model for logical reasoning.
According to results from the SWE-bench benchmark, which measures programming capability, o3 clearly outperformed all models in its category, including "o3-mini" and "Claude 3.7 Sonnet."
"o3" achieved a performance score of 69.1%, while "o4-mini" reached a close 68.1%.
Among the standout features of the new models is their ability to "reason using images." They can read and analyze images, even those that are blurry or low-quality.
Through this feature, users can upload drawings, diagrams, or pages from PDF files, and the models will visually analyze their content within the reasoning chain to enhance answer accuracy.
Notably, and for the first time in OpenAI models, o3 can independently use all ChatGPT tools, including browsing, running Python code, understanding images, and generating them.
This capability, in particular, may signal a shift towards more "agentic" AI ā meaning the model can execute a series of tasks without direct user intervention. This was demonstrated in a demo where "o3" analyzed a scientific research poster, gathered information from the internet, and derived a conclusion not mentioned in the original text
Capabilities and Features of o3 and o4-mini.
The new models are currently available to subscribers of ChatGPT Plus, Pro, and Team plans. An enhanced version of "o4-mini," known as "o4-mini-high," is also available, offering greater accuracy by dedicating more time to processing responses.
Furthermore, an "o3-pro" version is expected to launch in the coming weeks, exclusively for Pro plan subscribers. This version utilizes greater computing resources to deliver higher performance.
For developers, the company will make the new models available via API with competitive pricing: $10 per million input tokens and $40 per million output tokens for the "o3" model.
"o4-mini," meanwhile, is priced the same as "o3-mini" at $1.10 per million input tokens and $4.40 per million output tokens.
Codex CLI Tool
In an effort to enhance local AI capabilities, OpenAI has launched the "Codex CLI" tool.
This tool functions as an open-source programming agent that can be run locally in a terminal environment, connecting the new models to the user's files and specific systems.
Codex CLI supports both the "o3" and "o4-mini" models, with support for GPT-4.1 planned for later.
Separately, the company has launched a $1 million funding initiative to support experimental projects using these models, offering grants of up to $25,000 per project.
Finally, according to Sam Altman, "o3" and "o4-mini" are likely to be the company's last models in this reasoning series.
Therefore, it's clear that all these steps are fundamental groundwork leading up to the launch of GPT-5 ā the anticipated model expected to combine the capabilities of traditional models and reasoning models into a unified system.
