Skip to content

The agent engineering platform.

fix(langchain): use usage metadata scaling in SummarizationMiddleware default token counter (#35001)

ccurmef720ad00ae90
diff --git a/libs/langchain_v1/langchain/agents/middleware/summarization.py b/libs/langchain_v1/langchain/agents/middleware/summarization.py
index 8e9edca3e6..ab36cd554c 100644
--- a/libs/langchain_v1/langchain/agents/middleware/summarization.py
+++ b/libs/langchain_v1/langchain/agents/middleware/summarization.py
@@ -144,8 +144,10 @@ def _get_approximate_token_counter(model: BaseChatModel) -> TokenCounter:
if model._llm_type == "anthropic-chat": # noqa: SLF001
# 3.3 was estimated in an offline experiment, comparing with Claude's token-counting
# API: https://platform.claude.com/docs/en/build-with-claude/token-counting
- return partial(count_tokens_approximately, chars_per_token=3.3)
- return count_tokens_approximately
+ return partial(
+ count_tokens_approximately, chars_per_token=3.3, use_usage_metadata_scaling=True
+ )
+ return partial(count_tokens_approximately, use_usage_metadata_scaling=True)
class SummarizationMiddleware(AgentMiddleware):