o1

OpenAI o1 is a reasoning-focused model series released on September 13, 2024, built on reinforcement learning rather than the GPT architecture — OpenAI's first product of its "Strawberry" project . Its core innovation is test-time compute scaling: the model spends seconds to minutes internally reasoning through problems via a hidden chain-of-thought process, testing approaches and correcting mistakes before answering . This yields dramatic gains in STEM domains — on the IMO qualifying exam, o1 scored 83% versus GPT-4o's 13% — though it lacks GPT-4o's browsing, file upload, and image processing capabilities . The release included o1-preview and a smaller o1-mini, with API access initially restricted to tier-5 developers . OpenAI researchers framed this as a paradigm shift from scaling training compute to scaling inference-time reasoning, with Greg Brockman calling it "System 2 thinking" unlocked through reinforcement learning-trained chain-of-thought .