OpenAI has announced two new AI models-o3 and o4-mini, which demonstrate improved capabilities in logic, maths, programming and image analysis.
The flagship o3 model is able to “pause for thought” before forming an answer, allowing it to solve complex problems more efficiently. It outperforms previous OpenAI models in tests in the areas of programming, science, image recognition and problem solving. In turn, the o4-mini is a lighter and more affordable version that combines high performance with speed and cost-effectiveness.
Both models are already available to ChatGPT users with Pro, Plus and Team subscriptions, and via APIs – Completions and Responses APIs.
Key features of the o3 and o4-mini models:
- understanding and analysing images (even blurry or low quality images);
- execution of Python code directly in the ChatGPT window (via Canvas);
- integrated web search for relevant queries;
- image processing: scaling, rotation, etc.;
- improved understanding of tables, charts, and formulas.
For tasks requiring maximum precision, OpenAI offers an o4-mini variant – high-on is slower, but spends more time on careful analysis and response generation.
Another, more powerful version is expected soon – o3-pro, focused on maximum quality, which will be available exclusively for ChatGPT Pro users.
Interestingly, OpenAI CEO Sam Altman initially planned to integrate o3 directly into GPT-5 without releasing it separately. However, active competition from Google, Anthropic, xAI and other players forced the company to change its strategy.
According to OpenAI, the o3 model makes 20% fewer errors than previous developments and has become a leader in Codeforces and SWE-bench tests.
It is possible that o3 and o4-mini are the last “standalone” models before GPT-5, which will combine language capabilities with logic and multimodality.