OpenAI Introduces o3 and o4-mini: AI Thinks Better, Sees More, and Acts Autonomously | OpenAI API | Chat OpenAI | OpenAI stock | Turtles AI

OpenAI introduces two new AI models: o3 and o4-mini. Both models demonstrate superior performance in complex reasoning, multimodal vision, and strategic use of built-in tools. They are now available on ChatGPT and via API.

Key Points:

• o3 is OpenAI’s most powerful model ever, excelling at structured reasoning and multimodality.
• o4-mini is lighter and more efficient, optimized for fast but complex tasks, with excellent performance for its cost.
• Both models can use all ChatGPT tools (web, Python, vision, image generation).
• They have also been enhanced in terms of security, thanks to new training data and LLM monitoring systems.

OpenAI officially introduced its new language models o3 and o4-mini, marking a significant evolution in generative AI. o3 represents the cutting edge in terms of computational capacity and analytical precision, while o4-mini stands out for its efficiency and reasoning speed, designed for high-volume tasks. Both are designed to overcome the limitations of their predecessors, offering greater accuracy, adaptability and multimodal understanding, enhanced by full access to all the tools integrated into the ChatGPT platform, such as web browsing, code interpreter, reading uploaded files and visual analysis.

The o3 model sets a new standard in academic benchmarks, including Codeforces and MMMU, also surpassing the previous o1 in key areas such as advanced mathematics, programming and understanding technical images. It demonstrates more structured reasoning and a dramatic reduction in critical errors, as well as a strong ability to evaluate and formulate hypotheses in multidisciplinary contexts. The o4-mini, on the other hand, stands out for its particularly advantageous performance-to-cost ratio, with peak performance in the AIME 2024 and 2025 tests, excelling in both STEM and non-STEM areas, such as data science.

Both models are able to use tools agentically: they not only apply them, but decide when it is strategically useful to use them, adapting their workflow based on the end goal. They can, for example, chain together web searches, write code, generate graphs and explain the results visually, all within a single session, and often in less than a minute. This ability makes them useful for complex tasks that combine numerical analysis, visual interpretation and information synthesis, such as forecasting economic trends, technical review or conceptual design.

A distinctive aspect is the deep integration of visual processing into the model’s reasoning: o3 and o4-mini not only interpret complex images but incorporate them into their “train of thought”. They can understand blurry photos, upside-down diagrams, low-resolution sketches and act on them — rotating them, enhancing them, analyzing them — as part of the problem-solving process. This ability opens the door to new applications in education, scientific research and visual engineering.

On the security side, both models were trained with a new suite of prompts to reject malicious or borderline content, particularly in areas such as biohazards, jailbreaks, and malicious code generation. Internal assessments confirmed that o3 and o4-mini met the updated Preparedness Framework’s security criteria, falling below the high-risk threshold in the areas monitored, namely cybersecurity, self-improvement, and biohazards.

With the launch of the new models, OpenAI also made available Codex CLI, an open-source tool designed to directly interact with the user’s terminal, combining visual and text inputs for high-performance computer-aided programming. In parallel, a $1 million funding program was announced, dedicated to projects that leverage Codex CLI and OpenAI’s APIs, with grants awarded in credits of up to $25,000 per proposal.

Access to the models is immediately available to ChatGPT Plus, Pro, and Team users, while free users can try o4-mini by selecting the “Think” option before submitting the request. Access for Enterprise and Education users is expected within a week. For developers, the models are already integrable via API, with support for advanced features such as reasoning tokens and internal tools integrated into the logic of the models.

With this new generation, OpenAI consolidates the convergence of natural dialogue, computational capabilities, and agentic tools, paving the way for models capable of not only conversing, but acting in real and complex contexts.