LangChain's langchain-openai 1.5.2 adds o-series token counting fix
New patch preserves reasoning boundaries and supports o-series models in token counts.
LangChain, the popular AI framework builder, shipped langchain-openai 1.5.2 on August 18. This patch release focuses on reliability and compatibility with OpenAI's newer model families. The most notable fix is support for o-series models (like o1 and o3) in get_num_tokens_from_messages. Previously, token counting could be inaccurate for these reasoning models, leading to unexpected truncation or billing errors. Now developers can correctly estimate token usage when building prompts and managing context windows.
Beyond that, the release preserves reasoning item boundaries—meaning chain-of-thought traces are kept intact when processing responses—and extracts gateway metadata from response headers, which helps with debugging and monitoring in production. The update also includes snapshot refreshes and a minor release preparation (1.5.2a1). It's a small but meaningful update for teams relying on LangChain to orchestrate OpenAI's reasoning models. Developers should upgrade to avoid token miscounts and take advantage of cleaner metadata handling.
- Get_num_tokens_from_messages now supports o-series reasoning models for accurate token counting
- Reasoning item boundaries are preserved during response processing to maintain chain-of-thought integrity
- Gateway metadata is now extracted from response headers, improving observability in production
Why It Matters
Accurate token counting and reasoning preservation are critical for cost control and reliable AI agent workflows.