Authorization: Bearer compose-... and keep using normal OpenAI-compatible HTTP.
Pick the right flow
What happens on a paid inference call
The important detail is the order. Compose does not guess the bill from the prompt. It prices the request from provider output: token usage for text and embeddings, media evidence for image/audio/video, and the model pricing stored in the generated catalog.Exact vs upto
Usage-priced inference uses
upto because the final token count or media evidence is known only after the provider responds.