A reasoning (or “thinking”) model is trained to write out its working before it gives an answer, often in a block you can expand or hide. That extra step makes it noticeably better at math, logic, planning and hard code, and it changes what the model costs you to run.
What it costs
- Tokens. The thinking can run from a few hundred to many thousands of tokens before the first word of the answer. At 10 tokens per second, 2,000 tokens of thinking is over three minutes (speed).
- Context. The thinking lives in the context window and grows the KV cache (why it matters). A long chat of reasoning turns fills it quickly.
- Speed matters more. For a reasoning model, “Runs well” is the practical minimum; “Slow” is painful (ratings).
When it is worth it
- Math, puzzles, multi-step planning, debugging and anything with a checkable answer.
- Agents that must decide what to do next.
When it is not
- Everyday chat, rewriting, summaries and simple lookups, where the answer is no better and you wait longer.
Many reasoning models can switch thinking off, or limit how long it goes on, through a setting or a prefix in the prompt. Where available, use it: think for the hard questions and answer plainly for the easy ones. Models tagged reasoning are listed in the model list, and this guide explains the tags.