Reasoning models "thinking models"
LLMs trained to produce long internal "thinking" traces before answering — better at maths, code, logic. Costs far more per query.
OpenAI released o1-preview in September 2024 — showed that test-time compute (letting the model "think" longer) gives the same quality jump as making the model bigger. Successors: o3, DeepSeek-R1 (open weights, January 2025), Claude with extended thinking, Gemini 2 Thinking, Grok 4 Heavy. Typically trained with reinforcement learning on problems where the correct answer is verifiable (maths, code). Costs 5–50× more than ordinary models per query but solves problems that were previously out of reach.