Are reasoning tokens billed?
Generally yes. Thinking tokens are usually billed as output tokens even when the full thinking is hidden or summarized, so high effort settings can make each reply noticeably more expensive.
A model trained to think through a problem in steps before answering, using more time and tokens for harder tasks.
A reasoning model is a language model trained to work through a problem before giving its final answer, producing intermediate steps, often called thinking, that may or may not be shown to you. Training rewards it for reaching correct answers on problems such as math, coding and logic, so it learns to break problems down, check its work and try alternatives. Many current models can switch this on or off, or let you set how much thinking effort to spend. More thinking usually means better answers on hard tasks, but also more time and more output tokens, which you pay for.
For complex coding work, such as planning a change across many files or tracking down a subtle bug, reasoning modes often help noticeably. For simple tasks, they mostly add cost and delay. Knowing this lets you match the effort setting to the task rather than paying for deep thinking on a one-line fix. Many coding agents expose this as an effort or thinking setting you can change per task.
Asked why a test fails only on Mondays, a reasoning model might spend a few thousand tokens thinking: check how dates are created, notice the test uses the current date, consider time zones, find that a week-start calculation is off by one, and only then answer. A quick non-reasoning reply might just suggest mocking the date without finding the bug.
The visible thinking is not a perfectly faithful log of how the model reached its answer. It is useful and often informative, but the model can state reasons that do not fully match what drove its output.
All 28 terms are in the AI glossary. For models side by side, see frontier vs open-weight.
Generally yes. Thinking tokens are usually billed as output tokens even when the full thinking is hidden or summarized, so high effort settings can make each reply noticeably more expensive.
No. Use higher effort for hard, multi-step problems and lower effort for routine edits and quick questions. Try a lower setting first and raise it if results are not good enough.
It is the older prompting version of the idea, asking a model to think step by step. Reasoning models are trained to do this themselves, usually more effectively, without being asked.