Streaming

Set "stream": true in the request body to enable streaming. The three endpoints use different event shapes:

Chat Completions

Data-only SSE with chat chunks, an optional usage tail, and a [DONE] terminator:

data: {"id":"...","choices":[{"delta":{"content":"He"}}],...}
data: {"id":"...","choices":[{"delta":{"content":"llo"}}],...}
data: [DONE]

Responses

Named SSE events (response.created, response.output_text.delta, response.completed, …), no [DONE]; the stream ends with a terminal Response event. On the bridged path a minimal supported event family is synthesized, sequence_number increases monotonically, and token usage appears only in the terminal event.

Messages (Anthropic-compatible)

Anthropic named events ending with message_stop:

event: message_start
data: {...}

event: content_block_delta
data: {"delta":{"type":"text_delta","text":"Hello"}}

event: message_stop
data: {}

Interruption semantics

• Failure before the first client-visible event: a normal protocol error is returned; the platform may already have retried across same-protocol candidates.
• Failure after the first client-visible event: the HTTP status no longer changes; the stream ends with a protocol-native error event and a successful terminal state is never faked.
• If you did not receive [DONE] / message_stop / a terminal Response event, treat the stream as incomplete.

Python example

stream = client.chat.completions.create(
    model="gpt-5.6-sol",
    messages=[{"role": "user", "content": "Hello"}],
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)