Streaming and Non-Streaming Responses

This page explains response handling for Claude native calls (/v1/messages); integration methods are covered in Claude Native Call Basics.

Non-Streaming Responses

{
  "id": "msg_xxx",
  "type": "message",
  "role": "assistant",
  "content": [{ "type": "text", "text": "……" }],
  "stop_reason": "end_turn",
  "usage": { "input_tokens": 24, "output_tokens": 180 }
}

• The main content is in the content array, organized by blocks (text / tool_use / thinking), and should be iterated block by block before concatenation;
• stop_reason: end_turn indicates normal completion, max_tokens indicates truncation, and tool_use requests a tool call.

Streaming Responses (SSE)

Add "stream": true to the request, and events are delivered by type:

Event	Meaning
message_start	Session starts, includes input token count
content_block_start	A content block starts (text/tool_use/thinking)
content_block_delta	Incremental content; concatenate delta.text directly
content_block_stop	Current content block ends
message_delta	Trailing metadata (stop_reason, output token count)
message_stop	Everything ends

with client.messages.stream(
    model="claude-sonnet-5", max_tokens=1024,
    messages=[{"role": "user", "content": "Write a short poem about the sea"}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Notes

• When using the official Anthropic SDK, these events are already encapsulated; only handwritten SSE needs event-by-event handling;
• Token usage is based on message_start (input) + message_delta (output) and matches billing;
• For parsing the thinking block in thinking mode, see Thinking Mode.