LLM-Anthropic 0.27 moves to the new standard, OpenAI followed suit two weeks ago

Changing dependencies in the Anthropic Python library

Simon Willison has released a new version of his llm-anthropic plugin (0.27). The main reason for the release is to ensure compatibility with the anthropic v1.0.0 Python library. It brings a major technical change: dropping the httpx library and moving to the newer httpx2 standard. OpenAI made the same move with their v3.0.0 library about fourteen days earlier. Willison used the Claude model itself (Fable 5 in Claude Code) to perform the update, feeding it the official migration guide.

For plugin developers, this means the need to update the ecosystem

The ecosystem around LLMs is evolving rapidly, and with it, the underlying networking libraries. The move to httpx2 by Anthropic (and previously OpenAI) means that maintainers of third-party tools and plugins must react quickly to preserve functionality. Willison's approach shows that the process of updating code to a new API version can be effectively automated using LLM assistants, provided they have good migration documentation.

Relying on LLMs to write code has its limits

While Willison successfully used Claude Code to perform the update, the give the model the documentation and let it rewrite it approach works well for straightforward migrations. If this were a more complex architectural change, a simple prompt to read a Markdown file would likely fail or introduce hidden bugs. Verification against existing tests is an absolute necessity in such cases.

The next signal will be the stabilization of the network stack

The true measure of ecosystem maturity will not be the speed at which LLM vendors change their dependencies, but how stable their SDKs remain over the longer term. We should watch whether the move to httpx2 actually delivers better performance and resolves issues with asynchronous requests, or if it is just a temporary trend that will require another rewrite next year.

Lilith's verdict

Writing migration code via prompts is great for demos. Once we deploy critical infrastructure on it, we'll wake up in a hell of broken dependencies.