Openai Today he launched two breakthrough AI models that can justify images and use tools independently, representing what experts call a step in the possibilities of artificial intelligence.
The company based in San Francisco introduced O3 AND O4-miniThe latest of “O-Seria” of reasoning models, which, he claims, are so far the most intelligent and talented models. These systems can integrate images directly with the reasoning process, search the network, activate the code, analyze files, and even generate images in one task flow.
“There are some models that seem a qualitative step into the future. GPT-4 was one of them. Today there will also be one of those days,” said Greg Brockman, president of Opeli, during a press conference announcing the edition. “These are the first models in which the best scientists tell us that they produce legally good and useful new ideas.”
Like recent Openai models “Think with images” to rework visual problem solving
The most striking feature of those recent models is their ability to “Think with paintings“-Not only them, but manipulate and reason about them as a part of the problem solving process.
“They don’t see just a picture – they think with him,” Opeli said in a statement sent to Venturebeat. “This unlocks a new class solving class that combines visual and text reasoning.”
During the demonstration at a press conference, the researcher showed how O3 can analyze the physics poster from the decade of the internship, independently move with complex schemes, and even specifies that the outcome was not present on the poster itself.
“You know what was simply read, at least like 10 different articles in a few seconds,” said Brandon McKenzie, a researcher at Openai working on multimodal reasoning. He estimated that the task would take him “many days only for me, to even like on board, return to my project, and then a few days to search literature.”
The ability to artificial intelligence to govern images in the technique of reasoning – increasing details, rotation of diagrams or pruning unnecessary elements – is a recent approach that, based on industry analysts, can revolutionize the fields of scientific research to education.
I had early access, O3 is an impressive model, it seems very talented. Some funny examples:
– Ethan Mollick (@emollick) April 16, 2025
1) broke the business justification I use in my class
2) Creating some SVG (images created by the CODE)
3) writing a limited history of two blocking gyres
4) hard science fiction space battle. pic.twitter.com/tk4pkvknot
Opeli management emphasized that these editions are represented not only by improved models – these are complete AI systems that can independently use and mix many tools when solving problems.
“We trained them to use tools by learning to strengthen – creating them not only in the use of tools, but to reasoning when to use them,” explained the company in the edition.
Greg Brockman emphasized the wide possibilities of using the models: “In fact, they use these tools in a chain of thoughts when they try to solve a difficult problem. For example, we saw how O3 uses 600 connections of tools in a row, trying to solve a really difficult task.”
This ability allows models to perform complex, multi -stage flows of labor without a constant human direction. For example, if it is asked about future energy consumption patterns in California, artificial intelligence can search for utility data on the Internet, save Python code to investigate them, generate visualization and create a comprehensive report – every thing as a single fluid process.
Openai is ahead of competitors with record efficiency on AI key test tests
OPENAI claims that O3 establishes recent latest reference points in key measures of AI’s ability, including CodeIN Swo-BenchAND Mmwanness. In the opinion of external experts, O3 apparently makes 20 percent of great mistakes than its predecessor of inauspicious tasks in the real world.
The smaller O4-Mini model is optimized in terms of speed and cost performance while maintaining strong reasoning. On Aime 2025 Mathematics CompetitionO4-Mini obtained 99.5 percent when he received access to Python’s translator.
“I really believe that thanks to this set of models, O3 and O4-Mini, we will see more progress,” said Mark Chen, head of Openai research during a press conference.
The time of this release is significant, coming only two days after OpenAi presented GPT-4.1which leads in coding tasks. The quick order of ads signals acceleration in the competitive AI landscape, in which Opennai faces the growing pressure from Google Twins modelsAnthropics Claudeand Elon Musk’s XAI.
Last month, Opeli closed, which is the largest private round of technology financing in history, Collecting $ 40 billion with a valuation price $ 300 billion. The company is reportedly considering building its own Social networkPotentially compete with the X Elona Musk platform and secure the reserved source of coaching data.
O3 and O4-Mini are superb in coding, so we release a recent product, Codex Cli to facilitate their use.
It is an agent encoding on your computer. Is fully open source and is now available; We expect it to enhance quickly.
– Altman itself (@sama) April 16, 2025
Like recent OPENAI models, they transform software engineering with unprecedented code navigation skills
Software engineering is one of the areas where recent models, especially Excel, is software engineering. During the press conference, Brockman noticed that O3 is “actually better than me in navigation by our OpenAI code base, which is really useful.”
As a part of the commercial, he also introduced OpenAI Codex CliA light-weight coding agent who operates directly in the user terminal. The Open Source tool allows programmers to use the possibilities of reasoning models to coding tasks, with screenshots and sketches.
“We also share a new experiment: Codex CLI, a light coding agent that you can start from the terminal,” the company announced. “You can get the benefits of multimodal reasoning from the command line, transferring to the model screenshots or sketches with low loyalty, combined with access to the code locally.”
To encourage adoption, OPENAI introduces Initiative $ 1 million To support projects using Codex Cli and OpenAI models, with subsidies for increments of USD 25,000 in API loans.
Inside OpenAi’s Enhanced Safety Protocols: How the company protects against improper use of AI
Opeli reports on wide security tests of recent models, especially focusing on their ability to refuse harmful conclusions. The company’s security measures completely include the reconstruction of knowledge on security training and the development of the system to be soothed to mark dangerous hints.
“We tested both models with our most stringent security program,” said the company, noting that each O3 and O4-Mini remain below the “high” OpenAI threshold for potential risks in the field of biological, cyber and AI.
During the press conference, scientists Opeli Wenda and Anaana presented detailed comparative results, noting that the recent models have undergone over 10 times the training calculations of previous versions to realize their capabilities.
When and how will you access O3 and O4-Mini: Implementation Schedule and business strategy
The recent models are available immediately Chatgpt plusIN ProfessionalAND Team Users from Undertaking AND Education Customers gain access next week. Free users can try O4-Mini by selecting “Think” in the composer before sending inquiries.
Developers can access each models via OpenAI’s API CHECTION AND API answersAlthough some organizations will need verification to access them.
The edition is a significant business opportunity for Opeli, because the models seem each more capable and more profitable than their predecessors. “For example, in the AIME 2025 mathematical competition, the performance limit of performance for O3 is strictly improving compared to O1 and similarly the O4-Mini border is strictly improving in relation to O3-Mini,” the company said.
The way forward for artificial intelligence: how OpenAi fills in reasoning and conversation for recent generation systems
Industry analysts perceive these editions as a a part of a broader convergence of AI’s capabilities, with models increasingly combining specialized reasoning with natural conversation skills and the use of tools.
“Today’s updates reflect the direction in which our models are heading: specialized possibilities of reasoning of the series with a greater number of natural conversation skills and the use of the GPT series are converged,” Openai noted in its edition.
Ethan Mollick, an associate professor at Wharton School Who Studies at Adoption, described O3 as “a very strong model, but still jagged” in a post in social media after announcement.
Because competition in AI space is still intensifying, along with Google, anthropic and others releasing more and more stronger models, the OpenAi double emphasis on the possibilities of reasoning, in addition to the practical use of tools suggests a strategy aimed at maintaining its managerial position by ensuring each intelligence and utility.
With O3 and O4-Mini, Opennai crossed the threshold in which the machines begin to perceive images as people do-manipulative visual information as an integral a part of the pondering process, and not only analyzing what they see. This transition from passive recognition to lively visual reasoning can ultimately be more significant than any test result, representing the moment when AI began to see the world with pondering eyes.