Streaming (SSE)
Responses stream over Server-Sent Events. In chat, "stream": true delivers each chunk as data: {...} and the stream ends with data: [DONE]; the final chunk before that carries the authoritative cost and token counts. The Responses endpoint emits named events, and predictions have their own stream at GET /v1/predictions/{id}/stream.
Chat event format
Each data: line is one standard OpenAI chunk (choices[0].delta.content). For slow models, non-streaming responses may carry the advisory X-Suggest-Stream: true header. Turn streaming on to avoid client-side timeouts.
Named Responses events
The POST /v1/responses stream emits response.created, response.output_item.added, response.content_part.delta (text chunks), and a terminal response.completed carrying the usage block, then closes with data: [DONE].