o3-2025-04-16

The o3 model was officially released by OpenAI on April 16, 2025, and for the first time, it can autonomously utilize and integrate various tools from ChatGPT—such as web browsing, using Python to analyze uploaded files and other data, deeply understanding visual inputs, and even generating images. It can determine when and how to use these tools, producing detailed and well-structured responses swiftly (typically within one minute). The model excels particularly in tackling complex problems that require multi-level analysis and have less straightforward answers.

It can simultaneously process text, images, audio, and other modalities, seamlessly integrating data across these diverse formats. Architecturally, OpenAI has implemented an innovative neural network architecture that enables the model to encode image and text data into unified feature representations.

The model performs exceptionally well in visual tasks, significantly reducing major errors by 20% compared to OpenAI’s o1 when handling challenging real-world tasks. It demonstrates outstanding capabilities in areas such as programming, business/consulting services, and creative ideation.

Moreover, the model achieves top-tier performance across multiple domains, including programming, mathematics, and science. For instance, it showcases coding abilities comparable to those of elite programmers and problem-solving skills nearly on par with world-class mathematicians. In mathematics, it can accurately tackle highly difficult competition-level problems like those found in the AIME, achieving a success rate of 96.7%. In the scientific domain, it demonstrates remarkable accuracy—reaching 87.7% on the GPQA-Diamond benchmark for physics, chemistry, and biology expertise. In programming, its SWE-Bench Verified accuracy stands at 71.7%, surpassing both the o1 model's 48.9% and the o1 preview's 41.3%.

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