02 · Model Architectures
Reasoning Model
Trained to think before answering
A model optimized to spend additional inference compute on multi-step reasoning before returning a final answer. Some systems generate hidden reasoning tokens or use repeated sampling and search; extra compute can improve difficult math, logic, and coding tasks, but not uniformly.
Concrete example
Given a tricky logic puzzle, a reasoning model can spend more compute testing intermediate steps before it returns an answer.
Why it matters
Since 2024, test-time compute has become a major axis of model development alongside larger training runs.
Nearby in Model Architectures
Evidence
Sources for this definition
- 1Learning to reason with LLMs
OpenAI · Announcement · checked 2026-07-13