New models with visual reasoning from OpenAI: O3 and O4-Mini

OpenAI introduced two new models that pointed to an important leap in the field of artificial intelligence: O3 ve o4-mini. These models stand out as artificial intelligence systems that can “think”, not only with text, but also with visuals, can use vehicles on their own and complete multi -step tasks independently. The company states that these new generation models are the most competent systems they have developed to date in terms of both intelligence and practicality.

Human -like visual reasoning

OpenAI’s “visual thinking“The feature that he describes is the basic element that distinguishes O3 and O4-Mini from its similar. Users can now analyze and reason through a white-board photo, a graphics or hand-drawn diagram from PDF, and the models can reason.

OpenAI summarizes this ability by saying, “They don’t just look at the image; they think with it.” This paves the way for a brand new problem -solving class where visual and textual reasoning combines. Models can perform processes such as zooming, rotating and cropping on visuals independently.

Integrated with vehicles, multi -step task management

OpenAI emphasizes that O3 and O4-Mini are not only advanced models, but also full-fledged artificial intelligence systems. Models can chain all the vehicles in Chatgpt, including web browser, Python code manager, file analysis and visual manufacturer, and perform multi -step tasks from start to finish.

This feature allows the models to solve complex tasks without continuous orientation from the user. For example, the user said, “How will energy consumption in California change this summer compared to last year?” when he asks, The Model is looking for data from the web, writing Python code, creates a graphic, and explains its findings with a report.

Academic Achievements and Software Development

OpenAI’s O3 is particularly remarkable in the field of software development. OpenAI President Greg Brockman said that the model has found better direction than itself even at the code base of its companies. O3, Codeforces, SWE-Bench and MMMU set a new standard in leading tests, especially in areas such as software engineering, scientific research and business analytics.

O4-Mini, a smaller and cost-oriented model, makes a successful balance between speed and efficiency. 2025 AIME MATHEMATICS COMPETITION 99.5 percent success The providing model offers an ideal solution for high -volume uses.

NEW VEHICLE: Codex CLI and Support for Developers

Another innovation introduced with the models was the light terminal -based coding assistant called Codex CLI. Developers can take advantage of the coding capabilities of models with inputs such as screenshot or a sketchy drawing using this tool. Codex CLI was released as an open source via Github. In addition, the OpenAI launched a total grant program of $ 1 million in order to support projects using this vehicle.

Security and Ethical Tests

OpenAI announced that O3 and O4-Mini have been the most comprehensive security tests they have implemented so far. In particular, new rejection protocols were developed against biological threats, malware production and Jailbreak trials. The company says that the models are below the “high risk” threshold and have security measures against dangerous commands.

Use and pricing

O3 and O4-Mini Models, from today Chatgpt Accessible by Plus, Pro and Team users. Enterprise and Education plans will provide access within a week. Free users will be able to experience O4-Mini using the “Think” option at Composer.

Models for developers, as well, Chat Completiions started to be presented via API and Responses API. In the use of O3, the price of a million input token was determined as $ 10 and $ 40 per Output Token. For the smaller O4-Mini, these prices are $ 1.10 and $ 4,40, respectively.

OpenAI’s view of the future: Road to GPT-5

OpenAI CEO Sam Altman, these models can be the last independent reasoning models in Chatgpt, because the GPT-5 with reasoning capabilities and natural dialogue understanding announced. In this direction, the O3-pro model is on the road. A structure that combines the GPT-4.1’s coding power with the visual and logical intelligence of O3 is aimed.