OpenAI launched its latest o3-mini model, which is used in ChatGPT and API, refreshing benchmarks in areas such as mathematical code. o3-mini performs best in terms of accuracy and calibration error, and is 63% cheaper than o1-mini. Developers can choose different inference strengths to balance speed and accuracy. OpenAI co-founder Sam Altman said that o3-mini is a response to the DeepSeek craze, and revealed that the full-blooded version of o3 will be launched in the next few weeks. At the same time, he acknowledged past mistakes in open source strategy and is discussing a new open source strategy. Advanced speech model updates are also planned, possibly named GPT-5.