Anthropic Messages API — Streaming

Type: official-doc Tier: 1 (Official Doc) Author(s): Anthropic Date: Accessed 2026-06-01 URL: https://platform.claude.com/docs/en/docs/build-with-claude/streaming Accessed: 2026-06-01

Why This Source Matters

Defines how a response is delivered token-by-token instead of all at once. Streaming is what makes a chatbot feel responsive: the user sees text appear immediately rather than waiting for the full completion. This source gives the exact event protocol so the learner understands what is actually arriving over the wire, not just what the SDK abstracts away.

Key Claims

Enabling streaming

• Set "stream": true on a Messages request. The response is delivered as server-sent events (SSE).
• Each SSE has a named event type (event: message_stop) and a JSON data payload whose type matches.

Event flow (in order)

1. message_start — a Message object with empty content.
2. For each content block: a content_block_start, one or more content_block_delta events, then a content_block_stop. Each block has an index matching its position in the final content array.
3. One or more message_delta events — top-level changes to the final Message (e.g. stop_reason, final usage).
4. A final message_stop.

• ping events may be interleaved at any time; error events can also occur.

Content block delta types

• A text delta:

`` event: content_block_delta data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"ello frien"}} ``

• A tool-use delta uses input_json_delta with chunked partial_json.

SDK usage (Python)

The official SDK abstracts the SSE parsing:

import anthropic

client = anthropic.Anthropic()

with client.messages.stream(
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
    model="claude-opus-4-8",
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

• stream.text_stream yields only the text deltas.
• stream.get_final_message() returns the complete accumulated Message (identical to .create()), useful when you also need final usage.
• For large max_tokens, the SDKs require streaming to avoid HTTP timeouts.

Relevant To

• concepts: [streaming, server-sent-events, content-block-delta, token-by-token-generation]
• projects: [01-ai-chatbot]

Notes

• Final token usage arrives in the message_delta / final message, not in the first event — so cost can only be finalized after the stream completes.
• LiteLLM exposes streaming via stream=True, yielding OpenAI-style chunks (chunk.choices[0].delta.content) rather than raw Anthropic events.