OpenAI o3-mini - OpenAI's new inference model

What is OpenAI o3-mini?

OpenAI o3-mini is a brand-new inference model officially released by OpenAI. It is the latest and most cost-effective model in the OpenAI inference series, optimized for scientific, mathematical, and programming technologies. It supports three inference intensities (low, medium, and high), allowing users to adjust them to balance speed and accuracy. At medium inference intensity, o3-mini performs comparably to o1, with faster response times. o3-mini is the first model to offer inference functionality to free users, supporting features such as function calls, structured output, and developer messages.

Main functions of OpenAI o3-mini

• Strong STEM reasoning ability
  • Mathematical abilityAt low reasoning intensity, o3-mini's mathematical ability is comparable to o1-mini; at medium reasoning intensity, it performs similarly to the o1 model; and at high reasoning intensity, o3-mini's mathematical reasoning ability even surpasses that of o1.
  • Scientific ProblemsIn the Doctoral Questions Test (GPQA Diamond), o3-mini performed well across different inference intensities, achieving a maximum accuracy of 79.7%, which is superior to o1.
  • Programming skillsIn programming competitions (such as Codeforces), the Elo score of o3-mini increases with the strength of inference, and at medium inference strength it is comparable to that of o1 model.
  • Software Engineering TestingIn the SWE-bench Verified test, the high-inference version of o3-mini achieved an accuracy of 49%, making it the best performing model among all currently released models.
• Flexible reasoning strength optionsThe o3-mini offers three inference intensity options: low, medium, and high, allowing users to choose the appropriate intensity based on their needs.
  • Basic version with low reasoning intensitySuitable for rapid response and scenarios with high speed requirements.
  • Inference strength in mediumThe default setting strikes a balance between speed and accuracy.
  • high reasoning strengthIt offers deeper thinking when dealing with complex problems, but its response time is slightly slower.
• Supports a variety of developer featureso3-mini is the first small inference model to support the following developer features:
  • function callYou can directly call preset functions to achieve more efficient task processing.
  • Structured outputTo make the information output by the model more organized, making it easier to analyze and apply.
  • Developer NewsProvides developers with more debugging and feedback tools.
  • StreamingSupports streaming output, suitable for real-time interaction.
• Search function integrationo3-mini integrates search functionality, providing up-to-date responses with links to relevant web sources.

The technical principles of OpenAI o3-mini

• Deep reasoning mechanismThe model simulates human thinking, progressively analyzing each level of a problem to arrive at the most reasonable answer. When faced with complex problems involving multiple fields, o3-mini first identifies the key elements of the problem and then conducts a comprehensive analysis based on existing knowledge. This ensures the model can examine the problem from multiple perspectives, avoiding the limitations of a single viewpoint.
• Optimized algorithms and structuresThe o3-mini reduces computational resource consumption while maintaining high accuracy through optimized algorithms and simplified structure. By reducing unnecessary parameters and complex calculation steps, the o3-mini model can run efficiently on lower-configuration hardware. According to test data, the o3-mini's response speed on mobile devices is nearly 30% faster than similar models, and its energy consumption is reduced by approximately 40%.
• Supports multiple inference strengthsThe o3-mini offers three inference intensity options (low, medium, and high), allowing users to adjust inference effort based on task complexity. This ensures developers can fine-tune model performance as needed, enabling them to "think deeper" while prioritizing speed when necessary.
• Reinforce learning and adaptive capabilitieso3-mini utilizes a large amount of high-quality datasets during training and continuously optimizes its reasoning capabilities through reinforcement learning algorithms. Its adaptive learning capability allows the model to adjust in real-time based on user feedback, continuously improving the accuracy of its responses.
• Security DesignThe o3-mini employs deliberative alignment, training the model to perform comprehensive reasoning on manually defined safety guidelines before responding to user prompts. This significantly outperforms GPT-4o.
• Multi-domain optimizationThe o3-mini is specifically optimized for the fields of Science, Technology, Engineering, and Mathematics, achieving a faster response time than its predecessor. For example, in the AIME 2024 mathematics competition and the GPQA Diamond test for doctoral-level scientific problems, the high-inference-strength version of the o3-mini outperformed its predecessor.

How to use OpenAI o3-mini

• Using o3-mini via ChatGPT
  • Free usersYou can experience the o3-mini by selecting the "Reason" function in the ChatGPT chat bar.
  • Paid users:
    • ChatGPT Plus and Team usersStarting February 1, 2025, o3-mini will be available, with its daily message limit increased from 50 to 150 messages compared to o1-mini.
    • ChatGPT Pro users: You can access o3-mini without restrictions.
    • ChatGPT Enterprise usersAccess will be granted in one week.
• Using o3-mini via APIo3-mini has been opened to some API users (applicable to users using tiers 3-5), and can be accessed.Official websiteDevelopers can access o3-mini via the following APIs:
  • Chat Completions API
  • Assistants API
  • Batch API
• Select reasoning strengtho3-mini allows developers to select low, medium, or high inference intensity to optimize model performance.

OpenAI o3-mini model pricing

• Enter Token
  • Standard Price$1.10 per million input tokens.
  • Cache discount price$0.55 per million input tokens.
• Output Token
  • Standard Price$4.40 per million output tokens.
  • Cache discount price$2.20 per million output tokens.
• Compared with other models
  • Compared to o1-miniThe input token price of o3-mini is 63% cheaper than that of o1-mini, and the output token price is 80% cheaper than that of o1-mini.
  • Compared to DeepSeek R1The input token price of DeepSeek R1 is $0.14 per million, and the output token price is $0.55 per million, which is significantly lower than that of o3-mini.

Application scenarios of OpenAI o3-mini

• Science and Mathematical Problem SolvingThe o3-mini excels in solving mathematical and scientific problems, making it suitable for handling complex math competition problems (such as AIME 2024) and doctoral-level scientific problems (such as GPQA Diamond). It provides more accurate and clearer answers, and its reasoning ability is significantly superior to previous models.
• Programming and Software Engineeringo3-mini performs exceptionally well 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 real-world programming problems.
• Education and learning supportThe o3-mini can be used as an educational tool to help students and teachers solve complex academic problems, providing detailed solution steps and logical reasoning processes.
• Enterprise applicationsThe o3-mini is suitable for enterprise-level applications requiring efficient inference and rapid response, such as data analytics and automated decision support systems. It can handle complex logical problems while maintaining low latency and high cost-effectiveness.
• Multilingual understandingThe o3-mini is capable of handling complex problems involving multiple languages, making it suitable for cross-language academic research and business applications.