Streaming

OpenAI API#

import osfrom openai import OpenAIclient = OpenAI(base_url="https://fullsendai.com/v1", api_key=os.environ["FULLSEND_API_KEY"])stream = client.chat.completions.create(    model="deepseek-v4.1-flash",    messages=[{"role": "user", "content": "Explain this stack trace: ..."}],    stream=True,)for chunk in stream:    if chunk.choices:  # the last chunk carries only usage        print(chunk.choices[0].delta.content or "", end="", flush=True)

Each data: line is a chat.completion.chunk. Text arrives in delta.content, a reasoning model's thinking in delta.reasoning_content, and tool calls as delta.tool_calls fragments to join by index (see Tool calling). The chunk with finish_reason is followed by one with usage and empty choices, then data: [DONE]. Full event example: API reference.

Anthropic API#

import osimport anthropicclient = anthropic.Anthropic(base_url="https://fullsendai.com", api_key=os.environ["FULLSEND_API_KEY"])with client.messages.stream(    model="deepseek-v4.1-flash",    max_tokens=1024,    messages=[{"role": "user", "content": "Explain this stack trace: ..."}],) as stream:    for text in stream.text_stream:        print(text, end="", flush=True)    final = stream.get_final_message()

Events follow Anthropic's order: message_start, ping, then content_block_start / content_block_delta / content_block_stop for each text or tool_use block, message_delta with the stop reason and token counts, and message_stop. Full example: API reference.

Usage and cost#

Usage always comes at the end: in the last chunk on the OpenAI API (whether or not you set stream_options.include_usage), and in message_delta on the Anthropic API. A stream is charged once it ends, for those counts.

Stopping early#

Close the connection (break out of the loop, call stream.abort() or controller.abort()) and the model stops too. You're charged only for what was generated; when the model hadn't sent its final count yet, output is estimated from the text at about four characters a token. The call shows as Stopped in Logs.

Errors mid-stream#

Problems found before the first token (a bad key, an empty balance, a busy model) come back as a normal JSON error with its HTTP status. If the model goes quiet for 60 seconds after the stream has started, it ends with an error event (upstream_timeout on the OpenAI API, an error event on the Anthropic API). Treat what arrived as partial. See Errors.

Behind a proxy#