Output tokens are what the model writes back: the answer, the code, a tool request, or text in a fixed format such as JSON. They are generated one at a time, which is why long answers take longer and why responses can be streamed as they are produced.
They usually cost more than input tokens, often several times more per token, because generating is more demanding than reading. In models that reason before answering, the intermediate thinking may count as output tokens and be billed, even if it is not displayed in full.
A maximum output length can be set. If it is too low, the answer can be cut off in the middle. If it is too high, you pay for rambling.
Ask for concise answers where they are enough, set a sensible limit, and watch output length in agents and reasoning settings, where it can grow quickly.