The race between artificial intelligence models becomes more and more tight: openai has just announced the new reasoning models O3 And O4-Minidesigned to reflect for a long time before responding.
There are two important news: a greater aptitude to operational autonomy (which makes them particularly suitable for functions agentics) and the ability to think with the images. Gemini, GrokMeta Ai, Deepseek and Claude are notified.
What are O3 and O4-Mini
Openai defines O3 his reasoning model More powerful, capable of passing previous models in the tests that measure the skills of mathematics, coding, reasoning, science and visual understanding.
Openai O4-Mini Instead, it is a smaller model optimized for rapid and economic reasoning: according to Openai it represents a competitive compromise between price, speed and performance.
O3 is particularly ferred in complex queries that require multifaceted analyzes and whose responses may not be immediately obvious, and works particularly well in visual activities such as The analysis of imagesgraphic and graphics.
In fact, we see it wearing in the SWE tests (which evaluates the ability to solve programming problems in Github), Aider Polyglot, Mathvista and Humanity’s Last Exam, to give some examples.
O4-Mini It is particularly performing instead in the Aime 2024 and 2025 tests (which evaluate the ability to solve mathematical problems), as well as overcome the predecessor on non -Stem activities and on domains as data science.
To give some number, in the test Swe-Bench (without custom shelf), O3 gets a score of 69.1%, and O4-Mini 68.1%, while Claude 3.7 Sonnet obtained 62.3%.
The door towards operational autonomy and the first to think with the images
But the important characteristics of these models go beyond the numbers. First of all, O3 and O4-Mini are the first models that can use in independent way All chatgpt tools, such as web navigation, Python, understanding and generation of images.
Openii makes an example: if he wonders how the energy consumption will be in an area this summer compared to last year, the model will be able to Search on the web the necessary data, write a Python code to generate a prediction, create a graphic or an image and explain the key factors behind the forecast.
The other novelty concerns the ability to “think” with images. This means that users can load images on chatgpt, such as splashes or pdf diagrams, and models will analyze the Images during their phase “Chain-Of-Thougt“Before answering.
Thanks to this new capacity, O3 and O4-Mini can interpret blurred images and low quality and can carry out activities such as zoom or rotate the images to understand them better (a bit like we saw the launch of Gemini 2.0).
In addition, the new models have been evaluated by the Safety Advisory Group (SAG) of Openai and do not reach the safety limit in biology, chemistry, computer security or al.
The CEO of Openai Sam Altman said that O3 and O4-Mini could be the latest independent reasoning models before the launch of GPT-5, a model that, as happened for GPT-4.1 for traditional models, will unify the reasoning models.
Who can use the new tools
Starting today, the models, as well as a variant of O4-Mini call “O4-Mini-High“Which dedicates more time to providing answers to improve its reliability, are available for subscribers to the Pro plans (with unlimited access), Plus and Team (both with a limit of 150 queries per day) of Openai.
In the coming weeks, Openai says it plans to release o3-proa version of O3 that uses more computer resources to produce its answers, exclusively for subscribers Chatgpt Pro.
In addition to Chatgpt, all three models (O3, O4-Mini and O4-Mini-High) will be available for developers through the chat completions API and the API Responses at the following prices:
- O3: $ 10 per million input tokens (about $ 750,000) and 40 dollars per million output token
- O4-Mini: $ 1.10 per million input tokens and $ 4.40 per million output tokens.
Openi is also launching Codex CLIa light, open-source programming agent, which works locally in the user’s terminal.
To support the first projects, we are also starting a 1 million dollar initiative. Funding will be assigned in 25,000 dollar tranches in the form of API credits.
