Streaming

Read the response incrementally

With stream: true, responses use Server-Sent Events (SSE). SDKs parse events for you; when reading HTTP directly, process by event boundaries — never assume each network chunk is complete JSON.

from openai import OpenAI

client = OpenAI(
    base_url="https://189.24.66.33/v1",
    api_key="sk-relay-...",
)

stream = client.chat.completions.create(
    model="glm-5.3-flash-xploded",
    messages=[{"role": "user", "content": "Hello"}],
    stream=True,
    stream_options={"include_usage": True},
)

for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="", flush=True)
    if chunk.usage:
        print("\nUsage:", chunk.usage)

Usage and the final event

OpenAI streaming asks the upstream for usage. The final chunk containing usage may have no choices — check them separately. The Anthropic API uses its own message event format.

After a disconnect, settlement uses the usage actually received; token costs are never estimated when usage is missing. Check line-item results in usage records.