Why Reasoning Models Changed Everything

OpenAI released o1 in September 2024, and DeepSeek released R1 in January 2025.

Kaplan et al. reported that language model performance follows a power law with training compute.

In 2022, Wei et al. showed that step-by-step reasoning improved math and logic tasks.

Chain-of-thought prompting has been around since 2022, and reasoning tokens can act as a scratchpad.

OpenAI’s o1 system card says the model was trained with reinforcement learning to produce a chain of thought before answering.