Output tokens
The pieces a model generates for its reply. They count toward cost and latency, and an output limit can cut the answer short.
See it
What it is
Output tokens are the chunks of text and other content a model generates after reading a request. The model produces them one at a time until it reaches a stop condition or an output limit. They are counted separately from input tokens in most API usage reports, and they usually carry the higher per-token price of the two.
Watch output tokens when answers are slow, expensive, or getting clipped. A shorter requested format usually saves more than asking for vague conciseness, and streaming only shows tokens sooner; it does not reduce how many the model generates.
Gotcha: a token is not a word, and the maximum output setting is a ceiling, not a target. A response can stop earlier because it finished, hit a stop sequence, called a tool, or was blocked. Read the API's finish status and usage fields instead of estimating from the visible word count.
Ask AI for it
Instrument this OpenAI Responses API integration to record usage.output_tokens, model, latency, and response status for every call. Set max_output_tokens from configuration, surface incomplete_details when the limit truncates a response, and add daily totals grouped by model and endpoint. Keep input and output token counts separate, and alert when the 95th percentile output count exceeds 80 percent of the configured cap.