OpenAI Introduces o3 and o4-mini — Smart Models with Advanced Capabilities

OpenAI introduces new o3 and o4-mini models, the smartest to date. They can analyze queries, leverage all of ChatGPT’s tools, including web search, file analysis, visual data analysis, and image generation, and use Python for complex computations. These models automatically determine which tools to use and when to provide accurate and thoughtful answers in a minute or less, making them a powerful solution for complex problems.

OpenAI o3 is a flagship model with advanced capabilities in programming, mathematics, science, and image analysis. It sets new records in Codeforces, SWE-bench, and MMMU tests, demonstrating 20% ​​fewer errors than o1, especially in programming, business analysis, and creative tasks. The model copes well with hypothesis generation and testing in biology, engineering, and other sciences.

OpenAI o4-mini is a compact and cost-effective version optimized for fast computation. It outperforms previous models in math, programming, and visual tasks, and is also effective in data science. With its high speed and low cost, o4-mini supports large query volumes while remaining a powerful tool for everyday use.

Both models have become more natural in communication, take into account the context of previous dialogues and more often refer to reliable sources. They have been trained not only to use tools, but also to choose the best ways to solve problems.

Security is a key priority. o3 and o4-mini have reworked their training data, strengthening their protection against biorisks, malware, and jailbreaks. Internal tests show high resistance to dangerous queries. An additional monitor based on LLM, trained on human specifications, detects 99% of risks, making the models reliable to use.