DocumentationAPI reference

API reference

Responses as they happen.

Receive incremental output through Server-Sent Events.

Enable streaming

#

Set stream: true in a supported generation request. The gateway forwards events incrementally and keeps the selected upstream connection for the life of the stream.

Python · Responsespython
with client.responses.stream(
    model="gpt-5.6-terra",
    input="Write a short welcome message.",
) as stream:
    for event in stream:
        if event.type == "response.output_text.delta":
            print(event.delta, end="", flush=True)

Handle complete events

#

An SSE event ends with a blank line. Network chunks do not necessarily align with events or UTF-8 characters. Use an SDK or an incremental SSE parser rather than treating every received chunk as complete JSON.

  • Responses streams use response lifecycle events.
  • Chat Completions streams send data events and a [DONE] marker.
  • Messages streams use Anthropic event names, including content-block and message lifecycle events.
  • A stream can begin successfully and later report an error. Check its completion event and error events.

Timeouts, cancellation and retries

#

Closing the client connection cancels the upstream stream. Reported token usage is recorded. If final usage is missing, the credit hold remains pending reconciliation; cancellation does not guarantee a free request.

Do not blindly replay a started request

A connection failure after output begins does not prove the request was unprocessed. Automatically replaying generation may produce duplicate work or usage. The gateway does not replay streams after they begin.

Rolling updates allow existing streams to drain before an instance stops. Network, host or provider failures can still interrupt a connection; applications should handle incomplete output.