Preserve Usage Tokens in v3 Streaming

What it does

With langchain-core >= 1.4.8 (shipped the same day as langchain==1.3.10), token-usage metadata that previously vanished during streaming is now preserved in the final event payload. When you consume a stream chunk by chunk through the v3 streaming protocol, the aggregate usage counts are still available at the end.

Why it matters

In production LLM apps, token usage drives cost tracking, rate-limit management, and debugging. Before this fix, streaming paths gave you no reliable way to read the final token count without switching to the non-streaming invoke() path. Now you get both: a low-latency streaming UX and accurate usage reporting for billing, observability, and optimization.

Example

$ for event in llm.stream("Explain token usage in one sentence."):
    events.append(event)
    print(event.content, end="", flush=True)
final = events[-1]
print(final.usage)

LangChain is a framework for building applications powered by large language models.
— Usage: prompt_tokens=12, completion_tokens=15, total_tokens=27

$ for event in chat.stream_events([HumanMessage(content="What is LangChain?")], version="v3"):
    events.append(event)
    if event.event == "on_chat_model_stream":
        print(event.data.chunk.content, end="", flush=True)

LangChain is a framework for building context-aware applications...
— Final usage: prompt_tokens=9, completion_tokens=18, total_tokens=27

Common flags

stream_events(version="v3")

Yield v3 streaming events from a chain or LLM with usage preserved

on_chat_model_stream

Event type emitted per streaming chunk in the v3 protocol

final_event.usage

Usage metadata now present on the last event after streaming