OpenAI launches its most intelligent models to date: o3 and o4-mini with full access to tools

OpenAI has announced the launch of o3 and o4-mini, the latest models in its o-series characterized by their advanced reasoning capabilities. These models, described as the most intelligent released by the company to date, represent a significant change in ChatGPT's capabilities for both curious users and advanced researchers.

For the first time, these reasoning models can autonomously use and combine all tools available in ChatGPT, including web searches, file analysis with Python, visual reasoning, and image generation. The key innovation lies in their training to decide when and how to use these tools to produce detailed and thoughtful responses in appropriate formats, generally in less than a minute.

OpenAI o3 positions itself as the company's most powerful reasoning model, establishing new records in benchmarks such as Codeforces, SWE-bench, and MMMU. It is optimized for complex queries that require multifaceted analysis and particularly excels in visual tasks such as image analysis, graphs, and graphic elements. According to evaluations by external experts, o3 makes 20% fewer serious errors than its predecessor, OpenAI o1, in complex real-world tasks, excelling in programming, business consulting, and creative ideation.

Meanwhile, o4-mini is a smaller model optimized for fast and efficient reasoning. Despite its reduced size, it achieves remarkable performance in mathematics, programming, and visual tasks, being the best-performing model in the AIME 2024 and 2025 tests. In evaluations conducted by experts, o4-mini also outperforms its predecessor, o3-mini, in non-STEM tasks and domains such as data science.

One of the most notable innovations is these models' ability to integrate images directly into their chain of thought. They don't just "see" an image but "think" with it, unlocking a new class of problem-solving that combines visual and textual reasoning. Users can upload photos of whiteboards, textbook diagrams, or hand-drawn sketches, and the models can interpret them even if they are blurry or low quality.