Reasoning Model
Core ConceptsAI models specifically optimized for multi-step logical reasoning, math, and complex problem-solving — typically by using chain-of-thought at inference time.
Full Explanation
Reasoning models (OpenAI o1, Claude's extended thinking mode, Gemini Thinking) use 'test-time compute' — spending more tokens on internal reasoning before answering. They think through problems step by step internally, trading speed for accuracy on hard tasks. Best for math, coding, science, and multi-step logical problems. More expensive and slower than standard models.
OpenAI o1 uses extended reasoning to solve math competition problems that GPT-4o gets wrong.
Related Terms
A prompting technique where the AI is guided to reason step-by-step before producing a final answer, significantly improving accuracy on complex tasks.
A type of AI model trained on vast amounts of text data that can generate, summarize, translate, and reason about language.
A standardized test used to measure and compare AI model capabilities across specific tasks.