o1

OpenAI's o1 is a reasoning-focused large language model that marks a shift from pure pre-training scaling to inference-time scaling through reinforcement learning and chain-of-thought techniques . Unlike earlier GPT models that operate as "System 1" fast thinkers, o1 pauses to reason through problems—forming hypotheses, testing them, and backtracking when needed—taking 30 to 60 seconds per query in what OpenAI CPO Kevin Weil described as the "GPT-1 stage" of this new paradigm . The model introduces reasoning tokens during inference to expand compute budget at query time rather than just training time, achieving performance on par with PhD-level experts in physics, mathematics, and programming . Yichao Ji's experimental analysis of o1's API usage patterns suggested it likely performs linear autoregressive decoding rather than tree search, despite its extended reasoning behavior . The o1 release in September 2024 catalyzed industry-wide efforts to replicate its capabilities, including Moonshot AI's k1 visual reasoning model and open-source projects like Steiner, while subsequent o3 and anticipated o4 mini releases have extended the same trajectory toward longer inference times and higher capability ceilings .