TL;DR

OpenRouter now surfaces model reasoning (chain of thought) as standard content blocks, enabling display of a model’s thought process in applications like agent UIs.

Fixes Worth Knowing

• IBM integration now handles API errors during streaming responses.
• Google integration supports flex and priority pricing, and allows custom headers.
• OpenRouter CJS re-exports are now callable.

Before You Upgrade

If you are using OpenRouter with models capable of reasoning, update your application to handle the new reasoning content blocks in AIMessage objects to display the model’s chain of thought.

TL;DR

This release significantly improves support for structured outputs and reasoning traces, particularly when using OpenRouter and OpenAI models. The updates ensure reasoning content is surfaced correctly and tool call streams are handled more reliably.

Breaking

No breaking changes were introduced in this release.

New

• OpenRouter Reasoning: OpenRouter integrations now surface model reasoning as standard content blocks, allowing you to display the model's chain of thought.
• Anthropic Tool Calling: Anthropic now supports strict tool calling for custom tools.
• DeepSeek Models: Added support for DeepSeek V4 Flash and V4 Pro models.

Fixes Worth Knowing

• OpenAI Streaming: Fixed issues with streaming custom tool calls and preserving IDs in the Responses API.
• Agent Reliability: Improved agent stability by handling parsing failures in providerStrategy configurations more gracefully.
• JSON Parsing: Addressed potential errors when parsing JSON responses from OpenAI models.
• Bedrock Tool Calls: Correctly preserves Bedrock tool call identity in streamed chunks.

Before You Upgrade

No specific upgrade steps are required.

TL;DR

Gemini streaming now correctly includes usage metadata, improving observability when using Google’s generative AI models (large language models).

Fixes Worth Knowing

Streaming usage metadata is now correctly reported for Gemini models, allowing for better tracking and analysis of API usage.

TL;DR

LangChain.js now supports ToolRuntime (a way to define tools), improving tool integration and structured output parsing, especially with Ollama and OpenAI models.

Fixes Worth Knowing

• Message ordering after summarization is now correct.
• PDF loading now explicitly requires pdf-parse v1 (a PDF parsing library).
• Tool calls from Ollama models now correctly parse structured outputs.
• Chat models are now compatible with Vercel deployments (a web hosting platform).

Before You Upgrade

Ensure you are using pdf-parse v1 if you utilize the PDF loader.

TL;DR

LangChain.js now features more granular entry points (specific starting points) for increased modularity and flexibility.

New

LangChain.js now offers more granular entry points, allowing developers to import only the components they need, reducing bundle size and improving performance.

Fixes Worth Knowing

Prisma documentation has been updated for improved clarity and usability (database toolkit).