Thinking models
Also known as: reasoning models · extended thinking
Thinking (or reasoning) models are models trained to produce an explicit reasoning trace before the final answer — a scratchpad of intermediate steps.
They trade speed and tokens for accuracy:
- Slower — they generate many extra “thinking” tokens.
- Pricier — you pay for the thinking tokens too.
- Much better on hard math, logic, and planning.
When to use one: hard problems where accuracy beats speed. For simple lookups, a fast non-thinking model is usually the better (cheaper) choice.