o3-mini

o3-mini is the first lightweight intelligent model launched by OpenAI, featuring adjustable third-level reasoning capabilities. It’s a cost-effective, streamlined model in the reasoning series, specifically optimized for STEM fields while balancing speed and accuracy. The model offers three adjustable intensity levels—low, medium, and high—and, for the first time, grants free users access to its reasoning capabilities. At medium reasoning intensity, o3-mini delivers performance on par with o1, while at high intensity, it surpasses both o1 and o1-mini. In Codeforces, performance continues to improve as reasoning intensity increases. Additionally, in FrontierMath, o3-mini leverages Python tools to tackle 32% of problems for the first time.

In benchmark tests, o3-mini achieved an 83.6% accuracy rate at high reasoning intensity and 76.2% at medium intensity in the AIME 2024 competition; in the GPQA Diamond doctoral-level science test, it demonstrated 77.0% accuracy at high intensity and 71.5% at medium intensity. In the Codeforces programming competition, o3-mini scored an Elo rating of 2073 at high intensity and 1950 at medium intensity. In the SWE-bench Verified software engineering test, the model achieved 48.9% accuracy at high intensity and 42.3% at medium intensity.

o3-mini supports function calls, structured outputs, developer messages, batch APIs, and streaming output features, making it ready for direct deployment in production environments. It also introduces a new search functionality prototype that can locate up-to-date answers and provide links to relevant web resources.

Official reasoning model guide URL: https://platform.openai.com/docs/guides/reasoning?api-mode=responses