What are "reasoning" AI models?
“Reasoning” AI models1 are LLMs that spend some time and compute tokens “thinking” before answering queries. Examples include OpenAI’s o1 and o3, DeepSeek’s R1, Anthropic’s Claude Opus 4 and Sonnet 4, and Google DeepMind’s Gemini Flash Thinking.
These models use a process similar to chain-of-thought prompting to reflect on their own output. This allows them to perform better on tasks such as analytical reasoning.
Caption: DeepSeek’s R1 reflecting on a query
Reasoning models require substantially more compute than non-reasoning models, which increases their cost per output token.
As of Q1 2025, these models are rather new and there is no standardized name for this type of model. They have been called simulated reasoning models, chain-of-thought models, large reasoning models and enhanced reasoning models. ↩︎