langchain-anthropic==1.5.5 リリースlangchain-anthropic==1.5.5

The LangChain project has released langchain-anthropic 1.5.5, a patch update to the Python integration package that links the LangChain framework with Anthropic's Claude models. The release is narrow in scope, concentrating on bug fixes and data refreshes rather than new capabilities, but it is relevant to developers who monitor model usage or rely on Claude's tool-calling features inside agent workflows.

The headline change corrects how the library reports reasoning tokens in usage metadata. Recent Claude models can emit extended "thinking" or reasoning content alongside their final answers, and these tokens are typically counted and billed separately from ordinary output. Before this fix, that consumption appears to have been omitted or under-reported in the usage metadata that LangChain surfaces to applications. Accurate reporting matters because many teams use these figures for cost accounting, rate-limit management, and observability. Understated token counts can produce surprises on invoices or make it harder to reason about latency, throughput, and spend across a fleet of requests.

The second fix resolves a KeyError triggered by rename handling in the Claude file-tool middleware. Middleware in this context sits between the model and the tools it can invoke, transforming requests and responses as they pass through the chain. The file-tool middleware supports Claude's ability to work with files, and an unhandled rename operation could raise a KeyError that interrupted execution. Because such an exception can halt an agent mid-task, the correction is likely to improve reliability for pipelines that depend on file-based tools, particularly longer-running or multi-step agents where a single failure can abort the whole run.

Rounding out the release, the maintainers refreshed the model profile data. Model profiles generally describe the capabilities and constraints of each supported model, such as context window size, token limits, and which features are available. Keeping this data current helps the library route requests correctly and expose accurate metadata as Anthropic revises its model lineup. This kind of housekeeping change carries no behavioral risk on its own but keeps the integration aligned with the provider's latest offerings.

For readers less familiar with the layout, langchain-anthropic is one of several provider-specific integration packages that LangChain maintains outside its core library. In recent years the project has moved toward a modular architecture, separating the framework's core abstractions from partner integrations so that each provider package can be versioned and shipped independently. That structure lets a fix like this one reach Anthropic users quickly without forcing a broader release of the entire framework. It also means teams pinning specific versions should review these changelogs when upgrading, since bug fixes are distributed across many small packages rather than bundled into a single monolithic release.

The version number itself follows semantic versioning conventions. A move from 1.5.4 to 1.5.5 is a patch increment, signaling backward-compatible bug fixes rather than new features or breaking changes. In practice, most users should be able to upgrade without modifying their code, though testing is always advisable, especially for the usage-metadata change, which alters the numbers applications receive even if the underlying behavior of the model is unchanged.

The fixes also touch on themes that extend beyond this single package. Usage metadata feeds directly into observability and cost-tracking tooling, including LangChain's own LangSmith platform and similar third-party monitoring systems, so more accurate reasoning-token figures can ripple through dashboards and billing estimates. Likewise, file tools and middleware are increasingly common building blocks in agent frameworks, where models are given controlled access to documents and other resources; robust error handling in that layer is a recurring concern as agents take on more complex, stateful tasks.

Developers already running langchain-anthropic in production are the most likely to benefit, particularly those using Claude's extended reasoning modes or file tools. As with any patch release, the practical advice is to consult the referenced pull requests for exact details, verify that the reported token figures match expectations after upgrading, and confirm that file-based tool flows behave as intended in their own environments before rolling the update out widely.