OpenAI launches o3 and o4-mini, AI models that ‘think with images’ and use tools autonomously

Open Today, using two groundbreaking AI models, images and independent, using the means that experts have changed a step in a step in a step in the artificial intelligence, began something that can independently.

San Francisco-based company presented O3 and o4-miniThe latest in the most recent “in the” o-series “of the medical models he claims is the smartest and skilled models. These systems are able to integrate images into a direct motifs process, search, search the code, analyze the code and even analyze files.

“There are some models that feel like a quality step. GPT-4 was one of them. Today will be one of those days,” he said. “These are the first models that the best scientists say that they give good and useful novel ideas.”

https://www.youtube.com/watch?v=sq8gbpubpubpubpub3rk

‘Thinking with pictures” with images to solve the visual problem of Openai models

The most striking feature of these new models is their ability “Think with pictures“- They do not only see them, but treat them within the process of solving problems.

“They do not see only one image – they think of it,” he said. “This opens a class of a new problem that confuses visual and text justification.”

During a demonstration at a press conference, a researcher determined that the ten-year experiment could be able to manage a physical poster than ten years of experience, and even the posterior end of the final result.

“Readed, as 10 different documents in at least a few seconds,” Brandon McKenzie, in a few seconds in a few seconds working on multimodal justification during demo. The assignment “Only for me, for me, for me, and for me, and for me, for a few days, probably a few days to go through the literature.”

The ability to manipulate images in the AI’s reasoning process – to enlarge on the details, the diagrams can revolutionize the fields of scientific research areas of industrial analysts.

I experienced early access, O3 is an effective model, it looks very skilled. Some fun examples:
1) Breaking a job I used in my class
2) Create some SVGS (images created by code only)
3) Type a limited story of two interlocking gyres
4) HARD FANT FUND FIX WATCH. pic.twitter.com/tk4pkvnot

– Ethan Mollick (@emollick) April 16, 2025

Openai Executors stressed that these releases represent more than only improved models – full AI systems that can use the problems independently and combine multiple tools.

“We taught them to use tools with learning to strengthen them, teach you how to use the tools, but it has taught you to use them to think about how much you use them,” the company was released.

Greg Brockman, the models’ wide vehicle usage use: “In fact, they use these tools in the chain of thinking because they are trying to solve a difficult problem.

This ability allows models to complex, multi-step work flows without permanent human direction. For example, if you are asked about the rules for the use of future energy in California, EU can write a PYTHON code to search the Internet, analyze the Internet, analyze, create visual image and prepare a comprehensive report.

Openai is ahead of the ANI AI criteria with record-breaking performance

Openai claims O3 set new modern criteria among the main measures of the AI ​​ability Coders, Swenzaand Macwanness. In the assessments of foreign experts, O3, difficult, real-world positions, increasing by 20 percent of the predecessor.

The small o4-mini model is optimized for speed and cost efficiency while maintaining strong thinking. On Aime 2025 Math CompetitionWhen a Python translator was entered, O4-mini hit 99.5 percent 99.5 percent.

“Really, these sets, O3 and O4-mini,” Mark Chen, Openai’s research head Mark Chen, Markat Chen, brand Chen, branded chen, branded, then believe in brands.

This release is significant time, Openadi two days later after a few days GPT-4.1 modelsuperior in coding tasks. The fast sequence of ads, Openai signal a acceleration in a competitive AI view with an increase in pressure from Google TwinsAnthropical Claudand elon musk Xai.

Last month, Epenai closed what the largest private technological financing of history, Increases $ 40 billion In a $ 300 billion assessment. The company also states that it is considered to be established social networkPotentially compete with the X platform of Elon Musk and to provide the source of the owner of the owner.

O3 and O4-mini, super good in coding, so we release a new product to make it easier to use, the CODEX Cli.

This is a coding agent working on your computer. Completely open source and present today; We look forward to seeing it fast.

– Sam Altman (@sama) April 16, 2025

Openai’s new models transformed on the unprecedented code of software engineering

A field where new models are especially in Excel software engineering. Brockman said that during the press conference, O3’s Openai code is better than my side by passing through our database. “

As part of the announcement, Openai was also presented Codex cliA lightweight coding agent working directly in the user terminal. Support for open source tools, developers, screenshots and sketches, allows you to use modes of models for coding tasks.

“We share a new experiment: Codex Cli, a mild coding agent you can work from your terminal,” The company has announced. “You can get the benefits of multimodal justification combined with combined screenshots or low loyalty sketches, including images or code.

In order to encourage adoption, Openai is activated $ 1 million Initiative To support projects using Codex CLI and Openai models, there are grants in $ 25,000 in API loans.

Inside Openai’s advanced security protocols: How does the company protect the AI ​​abuse

Openai, the reports that have a wide security test in new models, especially harmful surveys are directed. The company’s security measures include security training information and the development of system-level eggs to build dangerous instructions.

“We have tested both models with our most serious security program.”

During the press conference, Wendai researchers Wenda and Ananya said that before the training score of previous versions to achieve the work of new models, the passage of new models.

When and how you can access O3 and O4-Mini-Mini: Placement schedule and commercial strategy

New models are available immediately Chatgpt Plus, Projectorand Team Users, Venture and Education Get access to customers next week. Free users can select the O4-mini example by selecting “Think” in the composer before submitting the questions.

Developers can get both models from Openai Chat complement the api and Answers APIAlthough some organizations need to be checked to access them.

Represents an important commercial opportunity for Openai because the graduation, models are more skilled and more economical than their predecessors. “For example, in the Math Mich race in 2025, the expense-performance border for O3 is strictly improving O1, and at the same time, O4-Mini’s borders are significantly improving O3-mini,” he said.

The future of the EU: How to Safety and Security for Result Systems

Industrial analysts are developing these releases within a wider convergence in AI capabilities, models, natural scratch skills and tools.

“Today’s updates reflect the management of our models: We combine specialized basic opportunities for the use of the GPT-series more natural scratch and tool, Openai, was released.

Ethan Mollick, Associate Professor in the School of Wharton, who learned to adopt AI, after the announcement, O3 described the “a very strong model, but still a torn”.

In the AI ​​space, it offers a strategy to protect the leadership of the competition, Google, Anthropic and Opens, using Openai’s double focus and a practical tool.

With O3 and O4-Mini, Openai, Openai, the images of the cars, passed through a threshold where he started to go through a threshold as an integral part of the thinking. This turn of passive recognition with active visual thinking can prove to be more important than each benchmark account that represents the world with EU’s thinking.