OpenAI releases o3 and o4-mini, new coding agent

The details: OpenAI o3 is the new top-tier reasoner, pushing SOTA performance across coding, math, science, and multimodal benchmarks. o4-mini offers fast, cost-efficient reasoning, significantly outperforming previous mini models and even saturating benchmarks like AIME 2025 math.

Both models can use and combine all tools within ChatGPT (web search, Python, image generation, etc.) as part of their problem-solving process. The models are also the first to be able to "think with images", integrating visual analysis and manipulation directly into their chain of thought. Also launching is Codex CLI, an open-source coding agent that runs in users’ terminals and links reasoning models with coding tasks. President Greg Brockman said the release is a “GPT-4 level qualitative step into the future,” with the models capable of producing novel scientific ideas. Why it matters: Whatever the bar for AGI is, it feels like the latest SOTA models are getting close. While reasoners were already a massive leap, equipping them with access to tools and multimodal capabilities has led to a class of models that is creating new ideas — seemingly taking us to Step 4 of OpenAI’s ladder of AI intelligence.