What is OpenAI o3-mini
OpenAI o3-mini is a new reasoning model officially released by OpenAI. It is the latest and most cost-effective model in OpenAI's reasoning series, specifically optimized for technical fields such as science, mathematics, and programming. It supports three reasoning effort levels (low, medium, high), allowing users to adjust according to their needs to balance speed and accuracy. Under medium reasoning effort, o3-mini's performance is comparable too1and it responds faster. o3-mini is the first to open up reasoning capabilities to free users, and supports features such as function calling, structured outputs, and developer messages.
Main features of OpenAI o3-mini
- Powerful STEM reasoning capabilities
- Mathematical ability: At low reasoning effort, o3-mini's mathematical ability is comparable to o1-mini; at medium reasoning effort, its performance is comparable to the o1 model; at high reasoning effort, o3-mini's mathematical reasoning ability even exceeds o1.
- Scientific problem handling: In the doctoral-level science question test (GPQA Diamond), o3-mini performs excellently at different reasoning effort levels, with a maximum accuracy of 79.7%, outperforming o1.
- Programming ability: In programming competitions (such as Codeforces), o3-mini's Elo rating increases as reasoning effort improves, and at medium reasoning effort it is already comparable to the o1 model.
- Software engineering tests: In the SWE-bench Verified test, the high-reasoning version of o3-mini achieved an accuracy of 49%, the strongest performance among all currently released models.
- Flexible reasoning effort options: o3-mini offers three reasoning effort options: low, medium, and high. Users can choose the appropriate effort level according to their needs:
- Basic low reasoning effort: Suitable for fast responses, for scenarios with high speed requirements.
- medium reasoning effort: Strikes a balance between speed and accuracy, and is the default setting.
- high reasoning effort: Provides deeper thinking when handling complex problems, but the response speed will be slightly slower.
- Supports multiple developer features: o3-mini is the first small reasoning model to support the following developer features:
- Function calling: Can directly call preset functions to achieve more efficient task handling.
- Structured outputs: Makes the information output by the model more structured and easier to parse and apply.
- Developer messages: Provides developers with more means of debugging and information feedback.
- Streaming: Supports streaming output, suitable for real-time interaction.
- Search feature integration: o3-mini integrates search functionality to provide the latest responses with links to relevant web sources.
Technical principles of OpenAI o3-mini
- Deep reasoning mechanism: The model simulates the way humans think, gradually analyzing all levels of a problem, and arriving at the most reasonable answer. When faced with complex problems involving multiple domains, o3-mini first identifies the key elements of the problem and conducts a comprehensive analysis in combination with its existing knowledge base. This ensures that the model can examine the problem from multiple perspectives and avoid the limitations brought by a single perspective.
- Optimized algorithms and structure: By optimizing algorithms and streamlining the structure, o3-mini reduces computational resource consumption while maintaining high accuracy. o3-mini reduces unnecessary parameters and complex computation steps, enabling the model to run efficiently on lower-spec hardware. According to test data, o3-mini's response speed on mobile devices is nearly 30% faster than similar models, with energy consumption reduced by about 40%.
- Supports multiple reasoning intensities: o3-mini offers three reasoning intensity options (low, medium, high), allowing users to adjust reasoning effort based on task complexity. This ensures developers can fine-tune model performance according to their needs, enabling it to "think more deeply" and also prioritize speed when needed.
- Reinforcement learning and adaptive capabilities: During training, o3-mini used a large amount of high-quality datasets and continuously optimized its reasoning capabilities through reinforcement learning algorithms. The adaptive learning capability enables the model to make real-time adjustments based on user feedback and continuously improve the accuracy of its responses.
- Safety design: o3-mini adopts deliberative alignment technology, training the model to fully reason about human-defined safety specifications before responding to user prompts. Significantly better than GPT-4o.
- Multi-domain optimization: o3-mini is specifically optimized for the fields of Science, Technology, Engineering, and Mathematics. It achieves faster response speeds than previous-generation models. For example, in the math competition (AIME 2024) and the doctoral-level science question test (GPQA Diamond), the high reasoning intensity version of o3-mini outperforms previous-generation models.
How to use OpenAI o3-mini
- Using o3-mini through ChatGPT
- Free users: You can do this in theChatGPTchat bar by selecting the "Reason" feature to experience o3-mini.
- Paid users：
- ChatGPT Plus and Team users: Starting February 1, 2025, o3-mini will be available, and its daily message limit will increase from o1-mini's 50 to 150.
- ChatGPT Pro users: Can access o3-mini without restrictions.
- ChatGPT Enterprise users: Will gain access one week later.
- Using o3-mini through the API: o3-mini has already been opened to some API users (applicable to users on tiers 3-5), and you can accessthe official websiteto check. Developers can use o3-mini through the following API:
- Chat Completions API
- Assistants API
- Batch API
- Selecting reasoning intensity: o3-mini supports developers in choosing from three reasoning intensities—low, medium, and high—according to their needs to optimize model performance
Model pricing for OpenAI o3-mini
- Input Tokens
- Standard price: $1.10 per million input tokens.
- Cached discounted price: $0.55 per million input tokens.
- Output Tokens
- Standard price: $4.40 per million output tokens.
- Cached discounted price: $2.20 per million output tokens.
- Comparison with other models
- Compared with o1-mini: o3-mini's input token price is 63% cheaper than o1-mini, and its output token price is 80% cheaper than o1-mini.
- Compared with DeepSeek R1：DeepSeek R1The input token price is $0.14 per million, and the output token price is $0.55 per million, significantly lower than o3-mini.
Application scenarios of OpenAI o3-mini
- Scientific and mathematical problem solving: o3-mini performs excellently in solving mathematical and scientific problems, and is suitable for handling complex math competition problems (such as AIME 2024) and doctoral-level scientific problems (such as GPQA Diamond). It can provide more accurate and clearer answers, and its reasoning ability is significantly better than previous-generation models.
- Programming and software engineering: o3-mini performs excellently in programming competitions (such as Codeforces) and software engineering tasks (such as SWE-bench Verified). It helps developers quickly generate code, optimize algorithms, and solve practical programming problems.
- Education and learning assistance: o3-mini can serve as an educational tool, helping students and teachers solve complex academic problems and providing detailed solution steps and logical reasoning processes.
- Enterprise-level applications: o3-mini is suitable for enterprise-level applications that require efficient reasoning and fast responses, such as data analysis, automated decision support systems, and so on. It can handle complex logical problems while maintaining low latency and high cost-effectiveness.
- Multilingual understanding: In multilingual understanding, o3-mini can handle complex problems in multiple languages, and is suitable for cross-language academic research and business applications.