OpenAI has introduced o3 and o4-mini: New flagship ChatGPT models with enhanced capabilities.

OpenAI has unveiled its latest artificial intelligence models - o3 and o4-mini. These models are updated versions of the 'o' series, focusing on extended reasoning before providing responses, setting new standards for performance and utility.According to OpenAI's announcement, the new models demonstrate significant progress in reasoning and integration with tools in ChatGPT. They can independently utilize and combine all ChatGPT tools, including file analysis using Python, in-depth visual data analysis, and image generation.These models can determine when and how to use tools to create detailed responses in the desired format, enabling complex problem-solving more quickly.Key features of the models:- OpenAI o3: The most powerful model, which has advanced capabilities in coding, mathematics, science, and visual perception. It has set new records on Codeforces, SWE-bench, and MMMU. The model is particularly effective for complex queries and visual tasks. It makes 20% fewer significant mistakes compared to the previous model o1 in complex real-world tasks, including programming, business consulting, and creative ideas.- OpenAI o4-mini: A smaller model optimized for fast and efficient reasoning. It shows good performance in mathematics, coding, and visual tasks. At AIME 2025, the model scored 99.5%, having access to the Python interpreter. It also surpasses the previous model o3-mini in non-STEM and data science tasks.Both models adhere to instructions and provide useful responses thanks to improved intelligence and integration of web sources. They have also become more natural in communication, using memory and conversation history to personalize responses.OpenAI stated that reinforcement learning (RL) enhances the performance of the models and has also incorporated images directly into the reasoning process. The models can analyze and manipulate images during reasoning.OpenAI o3 and o4-mini are not only smarter but also more cost-effective than the previous models o1 and o3-mini, respectively.OpenAI has also updated safety training data, developed a system for detecting dangerous queries, and launched the Codex CLI experiment. Users of various plans will have access to the new models, and there is also the opportunity to receive grants for projects using Codex CLI.The future of OpenAI lies in combining specialized reasoning capabilities with conversational skills and GPT-series tools to create even more powerful models.