PRODUCTION SCENARIO
Ostrava Legal's invoice shows 3.1 million output tokens for a contract-review job running gpt-5.6-sol at high reasoning effort. The answers returned to users add up to roughly 400,000 tokens.
Which token category accounts for the other 2.7 million output tokens?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Reasoning tokens, which are billed as output tokens but never returned to the caller
Reasoning models generate hidden reasoning tokens that occupy the context window and are billed as output tokens even though they are never returned; usage reports them under output_tokens_details. The other three are real costs but all land on the input side of the invoice: cached prefixes are discounted input, tool definitions are re-sent as input each turn, and a previous_response_id chain re-bills earlier turns as input.
Official Source: https://developers.openai.com/api/docs/guides/reasoning ↗