Also known as: thinking model
In plain English
A reasoning model "thinks" before it answers, working through a problem in steps like a person using scrap paper. It's slower, but better at maths, logic, planning and tricky questions.
In practice
Use reasoning models for complex analysis, coding and multi-step planning, and faster standard models for simple tasks like summaries or classification. The extra thinking uses more tokens, so it costs more and takes longer.
Under the hood
Reasoning models are trained, often with reinforcement learning on tasks with checkable answers, to generate an extended chain of thought before the final answer. Spending more compute at inference time, sometimes called test-time compute, improves accuracy on hard problems.
Example
"We switched to a reasoning model for the financial reconciliation checks."