OpenAI has released OpenAI o3 and o4-mini, the latest models in its o-series designed to think longer before responding. These models represent a step change in capabilities by being the first to agentically use and combine every tool within ChatGPT, including web search, Python code execution, and visual analysis.
- OpenAI o3 is the most powerful reasoning model, setting new state-of-the-art benchmarks on Codeforces, SWE-bench, and MMMU, while making 20 percent fewer major errors than OpenAI o1 on difficult tasks.
- OpenAI o4-mini is a smaller, cost-efficient model that achieves 99.5% pass@1 on AIME 2025 with a Python interpreter and outperforms its predecessor o3-mini in non-STEM domains.
- Both models integrate images directly into their chain of thought, allowing them to manipulate visual inputs like rotating or zooming as part of their reasoning process.
- The models are trained via reinforcement learning to reason about when and how to use tools, enabling multi-step workflows such as searching data, writing code, and generating graphs in under a minute.
These updates aim to provide significantly stronger performance across academic benchmarks and real-world tasks, establishing a new standard for intelligence and usefulness while often being more efficient than previous iterations.