OpenAI unveils o3 and o4-mini models with advanced reasoning and tool access

OpenAI has introduced two advanced reasoning models, o3 and o4-mini, which have achieved state-of-the-art results in AI benchmarks. These models are notable for having full access to external tools, including web browsing and a Python interpreter, marking a first for OpenAI.

The o3 model is particularly powerful, outperforming previous models in benchmarks like Codeforces, SWE-bench, and MMMU, and is capable of analyzing visual inputs through image uploads. Evaluations show that o3 reduces significant errors by 20% compared to its predecessor, o1, on complex tasks. The o4-mini model is designed for efficiency, optimized for high-volume reasoning tasks, and performs comparably to o3 across math, coding, and visual domains. It scored 99.5% on the AIME 2025 math benchmark when used with a Python interpreter. Both models are trained through reinforcement learning to effectively use tools and are described as more natural and conversational, with features supporting memory and prior context.

Available via ChatGPT Plus, Pro, and Team, these models replace previous versions, with Enterprise and Edu users gaining access soon. OpenAI also introduced Codex CLI, a new command-line tool that serves as a lightweight coding assistant developers can use directly on their local machines. Future updates will integrate additional tools like web search and code interpreter into the models' reasoning process.