o3-mini (high) is a specialized configuration of OpenAI's o3-mini reasoning model, optimized for complex STEM tasks including mathematics, science, and software engineering. The "high" designation refers to the maximum reasoning effort level, which allows the model to allocate more compute time and a greater number of reasoning tokens to its hidden chain-of-thought process before generating a final response. This configuration is intended for highly complex problems where accuracy is prioritized over latency.
Built as a more cost-effective successor to the o1 series, the model demonstrates significant improvements in logical deliberation and technical benchmarks. It utilizes reinforcement learning to refine its reasoning patterns, enabling it to perform at a level comparable to much larger models on specialized evaluations like the AIME 2024 and GPQA Diamond. Unlike earlier "mini" reasoning models from OpenAI, o3-mini supports advanced developer features such as function calling, structured outputs (JSON mode), and developer messages.
While the model is optimized for technical reasoning, it is designed with a smaller architectural footprint than the full o3 model to maintain efficiency. The high-effort setting specifically benefits multi-step coding projects and intricate scientific calculations by allowing the model to "think harder" and explore more potential solution paths during its internal processing phase.