Streaming
Per-protocol SSE start/complete, mid-stream errors, abort, retries, and where usage/cost appear. [DONE] is Chat Completions only.
Chat Completions#
SSE. The last event carries usage, plus cost once settlement has finished, and the stream ends with [DONE]. If cost is absent, look it up with X-Request-Id via GET /v1/generation. [DONE]
Messages#
Anthropic events. No [DONE]. X-Cost-* is omitted on stream.
Responses#
OpenAI Responses events such as response.completed. No [DONE]. Stream cost headers omitted. response.completed
Gemini#
alt=sse chunks. No [DONE]. Stream cost headers omitted. alt=sse
Retry & abort#
Provider fallback runs only before the first token. The server does not auto-retry image, video, or music generation requests. See error codes and retries & fallback for 429/503 retryable. errors · retries
Client abort#
If the client disconnects, only tokens already received are billed. Mid-stream upstream failures may arrive as SSE error events; see Errors during streaming for the frame format.