OpenAI o1 is a Reasoning model developed by OpenAI and released in preview form in September 2024, with a full release in December 2024. It represented a departure from the prior GPT-4 line by introducing a training and inference approach explicitly optimized for multi-step reasoning tasks such as competition mathematics, coding, and scientific problem-solving, rather than pursuing further general-purpose scale alone.
Extended chain-of-thought
o1's central technical innovation, as described by OpenAI, was training the model to generate an extended internal Chain-of-thought before producing a final answer, using Reinforcement learning to improve the quality of that intermediate reasoning process. Unlike earlier prompting techniques that elicited chain-of-thought reasoning through user-supplied instructions, o1 was trained to perform this reasoning automatically, and to allocate more computation to harder problems, an approach that became known as Test-time compute scaling: rather than only increasing model or training data size, capability could be improved by letting the model "think" longer at inference time. OpenAI deliberately hid the model's raw internal reasoning traces from users, showing only a summarized version, citing both competitive concerns about the technique being copied and safety concerns about the raw chain-of-thought containing unfiltered or manipulative content.
Performance and benchmarks
OpenAI reported substantial gains for o1 over GPT-4-class models on reasoning-heavy evaluations, including large improvements on competition mathematics problems and the SWE-bench software engineering benchmark, as well as strong performance on PhD-level science questions. The model's gains were most pronounced on tasks requiring extended, verifiable multi-step logic, while improvements on more open-ended or knowledge-recall tasks were comparatively modest, a pattern that reinforced the field's growing distinction between raw knowledge capacity and reasoning capability.
Industry impact
o1's release prompted rapid responses across the industry, with competing labs including DeepSeek releasing their own reasoning-focused models, most notably DeepSeek-R1 in January 2025, which achieved comparable reasoning benchmark performance at dramatically lower reported training cost and, unlike o1, published its full chain-of-thought reasoning traces and open weights. This contrast intensified debate over the value of OpenAI's decision to hide its reasoning traces and contributed to broader discussion of Open-weights models competition in frontier AI. o1 established test-time compute scaling as a distinct and durable paradigm alongside pretraining scale, with OpenAI and competitors subsequently releasing further reasoning-model iterations built on the same underlying approach.