OpenAI has unveiled its latest AI model, o3, which is more advanced than the previous o1 model announced in September. This model aims at providing enhanced answers to complex questions requiring detailed logical reasoning and outperforms o1 in various benchmarks, such as coding skills and math competency. Notably, o3 is three times better at tackling problems presented by ARC-AGI, a benchmark for testing AI's reasoning abilities in difficult math and logic problems. Meanwhile, Google has introduced a similar model, Gemini 2. 0 Flash Thinking, which excels in reasoning and agentic abilities. This development highlights the intense rivalry between OpenAI and Google, with both companies striving to push the boundaries of AI research.

OpenAI's CEO, Sam Altman, views this as a new phase in AI, enabling the execution of complex reasoning tasks. OpenAI's o3 model, also available in a smaller version called o3-mini, is not yet publicly available but is open for testing applications. The company has also shared more about the techniques used to enhance o1 with a method called deliberative alignment, which trains the model to evaluate requests and responses against safety specifications, making it harder to manipulate. The focus on logical reasoning is crucial as AI models are increasingly used for complex problem-solving on behalf of users. OpenAI and Google are racing to demonstrate their AI advancements, with both companies deploying models to handle diverse tasks like web browsing and smart assistant functions. Mark Chen, OpenAI's senior vice president of research, emphasized the utility advancement of the o3 model, while Altman praised its programming capabilities. Despite the absence of a major breakthrough at year's end, the frequency of AI announcements remains brisk, with OpenAI also introducing updates to its video-generating model, ChatGPT-powered search, and ChatGPT phone access.