OpenAI’s o3 and o4-mini Models Set the Standard for Model Performance and Efficiency

OpenAI is debuting OpenAI o3 and o4-mini, the latest additions to the company’s o-series of models which are designed to spend more time thinking before generating a response. As the smartest models OpenAI has released to date, o3 and o4-mini are engineered to use state-of-the-art reasoning and the wide variety of ChatGPT tools to respond to multi-faceted questions.

OpenAI o3 and o4-mini are designed to produce more thoughtful outputs, namely by employing an agentic method of determining, combining, and using the right ChatGPT tools to deliver a rich response. From searching the web to analyzing uploaded files, reasoning deeply about visual inputs, or generating images, OpenAI’s latest models are equipped with the ability to produce detailed responses while leveraging the appropriate tools to arrive at that response.

OpenAI o3 is a powerful reasoning model ideal for complex queries requiring multi-faceted analysis, especially in coding, math, science, and visual perception areas. The model excels at analyzing images, charts, and graphics, and particularly performs well in difficult, real-world tasks associated with programming, business/consulting, and creative ideation.

OpenAI o4-mini is a smaller model compared to o3, designed for fast, cost-efficient reasoning. Excelling in math, coding, and visual tasks, o4-mini outperforms o3-mini on both non-STEM and STEM tasks.

Both models are reportedly better at instruction following and generating useful, verifiable responses compared to their predecessors, according to OpenAI.

Other key capabilities of OpenAI’s newest models include:

• “Thinking” with images, integrating images directly into their chain of thought, blending visual and text reasoning for advanced problem-solving
• Full access to ChatGPT tools and custom tools via function calling the API, where the models are trained to reason about how to best solve problems with which tools and when
• Greater cost-performance efficiency, where OpenAI anticipates that o3 and o4-mini will be both smarter and cheaper than o1 and o3-mini, respectively
• Codex CLI, a lightweight coding agent that works directly on end-users’ computers, which maximize the reasoning capabilities of the o3 and o4-mini models
• Rebuilt safety training data and system-level mitigations that improve the safety of the models, including the addition of new refusal prompts and flagging dangerous prompts in frontier risk areas

To learn more about OpenAI o3 and o4-mini, please visit https://openai.com/.