llm-anthropic 0.25.1

The more consequential change here is not model support but the token default fix: previously, llm-anthropic imposed an 8,192-token ceiling across all models regardless of what those models could actually produce, meaning developers were silently getting truncated outputs without necessarily knowing the tool was the bottleneck rather than the model.

This update sits in a different part of the AI story than most of what Modelwire has covered recently. The Glean coverage from May 29 focused on enterprise buyers prioritizing cost containment over raw capability, and that framing does not map cleanly onto a developer-tooling patch. What this story actually belongs to is the quieter, ongoing work of making frontier model capabilities accessible at the command line and in scripts, where small defaults can silently constrain what practitioners think a model can do. The token ceiling issue is a good example of tooling lag: a model ships with expanded output capacity, but the wrapper around it preserves an older, more conservative assumption until someone notices and fixes it.

Watch whether other popular LLM wrappers (LangChain, LiteLLM) have analogous hardcoded token ceilings that have not been updated to match current model maximums. If several do, it suggests a systemic pattern where tooling assumptions quietly lag model capability across the broader developer stack.