OpenAI Unveils o3 and o4-mini

OpenAI has released o3 and o4-mini, the latest models in the o-series that think longer before responding. These models have the ability to actively use and combine all the tools within ChatGPT—including web search, data analysis using Python, visual reasoning, and image generation—like an agent.

The key features of the main models are as follows:

• OpenAI o3: The most powerful reasoning model, breaking through limits in coding, math, science, and visual perception. It achieved new SOTA results on major benchmarks including Codeforces, SWE-bench, and MMMU, and external expert evaluations showed 20% fewer major errors than OpenAI o1 on difficult real-world tasks.
• OpenAI o4-mini: A small model optimized for fast, cost-efficient reasoning. It shows excellent performance in math, coding, and visual tasks, achieving particularly outstanding results on the AIME 2024 and 2025 benchmarks. It also shows better performance than o3-mini in data science and non-STEM fields.

These models continuously scaled their performance through reinforcement learning (RL), and show improved instruction-following ability and more natural conversation compared to previous models. In particular, they remember past conversation content to provide more personalized responses to users.