OpenAI released the o1 model family between September and December 2024, introducing a new paradigm where models generate hidden reasoning steps during inference to solve complex problems in mathematics, science, and software engineering.
- o1-preview and o1-mini launched on September 12, 2024, utilizing reinforcement learning with process-supervised reward models to enable deliberation rather than immediate pattern matching.
- The full o1 model released on December 5, 2024, added vision input, function calling, and a 200K token context window, achieving a 34% reduction in major errors on difficult reasoning tasks compared to o1-preview.
- o1-pro, available via ChatGPT Pro subscription and later API access, extends the compute budget to increase AIME pass rates from approximately 74% to 86%.
These releases established what the industry calls the "reasoning era" by demonstrating that scaling inference-time compute unlocks capabilities beyond pretraining alone.