OpenAI o3-mini and o3-mini-high
OpenAI’s o3-mini and o3-mini-high models are fast, cost-effective solutions for STEM tasks. They are suited for working with code, mathematics, and data analysis.
OpenAI continues to improve artificial intelligence by introducing a new model, o3-mini — the most cost-effective and fastest model in its family of reasoning models. Released for use in ChatGPT and via the API, o3-mini is already available to developers and paid subscription users, while free users can try it through a dedicated “Reason” button.
Expanded capabilities and flexibility
One of the key innovations is support for three reasoning modes: low, medium, and high.
Low provides basic operation with minimal time costs.
Medium offers a balance between speed and accuracy, making it suitable for most everyday tasks.
High is aimed at handling complex tasks that require maximum accuracy and depth of reasoning, although this comes at the cost of slightly higher response latency.
This flexibility allows developers to choose the optimal mode for specific use cases, whether programming tasks, mathematics, or scientific questions. In addition, the model supports features such as function calling, structured outputs, and developer messages, making it “production-ready” from day one.
Optimized for STEM tasks
One of the priorities in developing o3-mini was high efficiency in STEM disciplines. Test scores show that the model demonstrates significant progress in solving math problems, PhD-level scientific questions, and competitive programming tasks.
Mathematics: when using high mode, accuracy reaches 83.6% on competition problems such as AIME 2024.
Science: scores on complex questions in biology, chemistry, and physics reach 77% accuracy.
Programming: in coding competitions such as Codeforces, the model shows a substantial increase in Elo rating, reaching 2073 points.
In addition, o3-mini demonstrates improved speed: responses are generated on average 24% faster, and time to first token has decreased by approximately 2500 ms compared with its predecessor, o1-mini.
Reliability and safety
Special attention has been paid to safety. The use of the deliberative alignment technique allows the model to first assess compliance with human safety requirements before responding to user requests. As a result, o3-mini successfully handles safety tasks and resists attempts to bypass restrictions (jailbreaks), as confirmed by comparative tests with other models, including GPT-4o.
Cost-effectiveness and scalability
A cost-effective approach to development not only reduces computing resource costs but also lowers the cost of processing each token. OpenAI has managed to reduce the cost of intelligence by 95% since the launch of GPT-4, making high-quality artificial intelligence more accessible. Paid subscription users receive expanded capabilities, including an increased message limit — from 50 to 150 messages per day for ChatGPT Plus and Team, while Pro users receive unlimited access.
Conclusions and outlook
The release of OpenAI o3-mini and its o3-mini-high variant demonstrates that even compact models can compete with larger counterparts in accuracy, speed, and functionality. Optimized for STEM tasks, these models are becoming indispensable tools for solving complex logical, mathematical, and programming problems while remaining cost-effective and fast.
The new models confirm OpenAI’s commitment to expanding the boundaries of artificial intelligence, making it available for a wide range of applications. Developers and users can already take advantage of o3-mini today, and even more ambitious projects combining high performance, safety, and efficiency await us in the future.
