Streaming

Streaming uses Server-Sent Events (SSE) to deliver tokens to the client as the model generates them. Instead of a 5-second wait followed by a wall of text, users see the response appear word by word, similar to watching someone type. This dramatically improves perceived latency and user experience. All major AI APIs support streaming: OpenAI (stream: true in Responses API), Anthropic (stream: true in Messages API), and Google (streamGenerateContent endpoint). On the client side, you process the SSE stream and append each token to the UI. Most AI chat interfaces use streaming by default. Streaming is essential for any user-facing AI feature. A 3-second time-to-first-token feels fast with streaming but painfully slow without it. Implementation considerations: you cannot parse structured JSON until the stream completes, error handling is different (errors may arrive mid-stream), and you need to handle connection drops gracefully. For batch/background tasks where no human is waiting, non-streaming is simpler.