Reasoning

Hidden or partially-visible step where the model thinks before answering. Anthropic exposes signed blocks; OpenAI bills opaque tokens.

Also known as: extended thinking

See also: thinking-block