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

Openai has launched two models of artificial intelligence today that can cause images and use tools independently, which represents what experts call a step to the abilities of artificial intelligence.

The San Francisco-based company, O3 and O4-MINI, which is the latest in the “O-Series” thinking models, which claims to be its most intelligent and capable models so far. These systems can integrate images directly into their thinking process, search in the web, play the code, analyze files, and even create images within one task flow.

“There are some models that you feel a qualitative step in the future. GPT-4 was one of these. Today it will also be one of those days,” Greg Brockman, president of Openai, said at a press conference announced. “These are the first models in which senior scholars tell us that they produce new and legally beneficial ideas.”

https://www.youtube.com/watch?

How to think of the new Openai models in the pictures to transform visual problems

The most exciting feature of these new models is its ability to “think about pictures”-not only to see them, but treatment and mind towards them as part of the problem solving process.

“They only see a picture – they are thinking about it,” Openai said in a statement sent to Venturebeat. “This opens a new category of solving problems that mix optical thinking.”

During a demonstration in the press conference, one of the researchers showed how O3 can analyze the physics poster from a period of time a decade ago, and transfer its complex plans independently, and even to determine that the final result was not present in the same poster.

“You must have just read, as you know, at least like 10 different papers in a few seconds for me,” said Brandon McKenzi, an Openai researcher who works on multimedia. It was estimated that the task would have taken it “for me only for me until I love, on the plane, and I go back to my project, then a few days, most likely, to search for literature already.”

The ability of artificial intelligence to process images in the thinking process – enlarging details, periodic plans, or cultivating unnecessary elements – is a new approach that industry analyst says can revolutionize areas from scientific research to education.

I had early access, O3 is an impressive model, and it seems very capable. Some fun examples:
1) The status of the work I use in my class
2) Create some svgs (photos created by software instructions alone)
3) Writing a bound story for two interlocking Gyres
4) The Battle of Space Fiction. pic.twitter.com/tk4pkvknot

Emollick April 16, 2025

Openai executive officials emphasized that these versions represent more than just improved models – they are fully artificial intelligence systems that can use multiple tools and link them independently when solving problems.

“We have trained them to use tools through reinforcement learning – not only how to use tools, but for the reason for the time they are used,” the company explained in its issuance.

Greg Brockman highlighted the possibilities of using wide tools for models: “They are already using these tools in their series of ideas because they are trying to solve a difficult problem. For example, we have seen O3 use like 600 consecutive calls in a row trying to solve a truly difficult task.”

This allows models to perform complex multi -step workflow tasks without a fixed human direction. For example, if asked about the patterns of future energy use in California, AI can search in the web to obtain the tool data, write the Python icon to analyze, create perceptions, and produce a comprehensive report – all as a single liquid process.

Openai progresses on competitors through record performance on the main artificial intelligence standards

Openai O3 claims new modern standards through the main measures of artificial intelligence, including Codeforce, Swe-Bench and MMMU. In the assessments conducted by external experts, O3 is said to make 20 percent less than the main errors of its predecessor in difficult and realistic tasks.

The smaller O4-MINI model is improved for speed and cost efficiency while maintaining strong thinking capabilities. In the AIME 2025 Mathematics Competition, O4-MINI recorded 99.5 percent when you can access the Peton translator.

“I really think it is with this wing of models, O3 and O4-Mini, we will see more developments,” Mark Chen, Openai’s head of research, said during the press conference.

The timing of this version is important, just two days after the OpenAI unveiled the GPT-4.1, which excels in coding tasks. The rapid caliphate of ads indicates an acceleration in the scene of competitive intelligence, where Openai faces increased pressure from Google Gemini, Clauds, and Elon Musk’s Xai.

Last month, Openai closed up to the largest special technical financing round in history, raising $ 40 billion with a rating of 300 billion dollars. According to the company, the company is considering building its social network, and it may compete with the Elon Musk X platform and to ensure a special source of training data.

O3 and O4-MINI are very good in coding, so we launch a new product, Codex Cli, to make it easier to use.

This is a coding agent that works on your computer. It is completely open and available today; We expect to improve quickly.

Sam Al -Taman (Sama) April 16, 2025

How to transform the new models of OpenAi software engineering with unprecedented navigation capabilities

One of the areas where new models are especially the Software engineering. Brockman noticed during the press conference that O3 “is really better than I am moving through the Openai code base, which is really useful.”

As part of the advertisement, Openai also Feel Codex Cli, a lightweight coding factor that works directly at the user station. An open source tool allows developers to take advantage of the possibilities of thinking about the coding tasks, while supporting screenshots and graphics.

“We also share a new experience: Codex Cli, a lightweight coding agent you can run from your station,” the company announced. “You can get the benefits of multimedia thinking from the command line by passing screenshots or low sincerity drawings of the model, as well as access to the code locally.”

To encourage adoption, Openai launches an initiative of one million dollars to support projects using Codex Cli and Openai models, with grants of $ 25,000 in API credits.

OpenAi safety protocols: How to protect the company from misuse of artificial intelligence

Openai reports that conduct widespread safety tests on new models, especially those that focus on their ability to reject harmful requests. The company’s safety measures include rebuilding full safety training data and developing a system of systematic levels for the mark of dangerous claims.

“We emphasize the test of both models through our strict safety program so far,” the company stated, noting that both O3 and O4-MINI remain below the “high” threshold for Openai for potential risks in cybersecurity and cybersecurity, and the capabilities of Amnesty International.

During the press conference, Openai Wenda and Ananya researchers presented detailed standard results, noting that the new models were undergoing more than 10 times the training account on previous versions to achieve their capabilities.

When and how you can access O3 and O4-MINI: The Table Publishing and Commercial Strategy

New models are immediately available to Chatgpt Plus, Pro and Team users, with access to institutions and education next week. Free users can take samples from the O4-MINI by choosing “Think” in the composer before submitting the queries.

Developers can access both models via the Openai and API API application programming interface, although some institutions will need to verify access.

The version represents an important commercial opportunity for Openai, as models appear more costly and more efficient in its ancestors. “For example, in the 2025 AIME mathematics competition, the cost-cost performance limits of O3 improve on O1, and likewise, the O4-MINI borders improved strictly on O3-MINI”, the company mentioned.

The future of artificial intelligence: How Openai bridges thinking and speaking to the following generation systems

Industry analysts view these versions as part of the broader rapprochement in the capabilities of artificial intelligence, as models are increasingly combining specialized thinking, natural conversation capabilities and tools.

“Today’s updates reflect the direction to which our models are heading: We are close to the specialized thinking capabilities of the O with more natural conversation capabilities and the use of tools for the GPT series,” Openai indicated in its release.

Ethan Malik, associate professor at the Warton School, who studies the adoption of artificial intelligence, described O3 as a “very strong model, but he is still registered” in a post on social media after the announcement.

With the continued competition in the area of ​​artificial intelligence in its intensification, with the presence of Google, the Anthropor, and other increasing models, Openai’s double focus on both thinking capabilities and the use of practical tools indicates a strategy aimed at maintaining its leadership position by providing both intelligence and interest.

With O3 and O4-MINI, Openai crossed the threshold where machines begin to perceive images the way humans do-processing visual information as an integral part of their thinking process rather than just analyzing what they see. This shift may be proven from negative recognition to active visual thinking in the end more important than any degree, which represents the moment when artificial intelligence began to see the world through eyes thinking.