llm-anthropic 0.26 adds support for Claude 5 models and server-side tools

Simon Willison’s llm-anthropic plugin version 0.26 adds support for Anthropic’s Claude 5 series—claude-fable-5, claude-sonnet-5, and claude-opus-5—via the underlying LLM 0.32 framework. The update also introduces server-side tools for WebSearch, WebFetch, CodeExecution, and AnthropicMCP, accessible through LLM’s -T interface or Python’s tools= parameter. Prior -o web_search* options have been removed in favor of -T WebSearch.

The release upgrades to LLM 0.32, which changes how reasoning, tool calls, tool results, and server-side tool results are streamed as typed events. Reasoning output for llm CLI prompts now writes to standard error by default unless the --hide-reasoning/-R flag is used.

Extended thinking controls have been simplified to the thinking and thinking_effort flags, which accept values of low, medium, high, xhigh, or max. The Claude 5 models enable thinking by default; users can disable it for Sonnet 5 and Opus 5 with -o thinking 0, while Fable 5 always performs thinking. The -R/--hide-reasoning flag now omits reasoning from both responses and logs. Prior options including thinking_budget, thinking_display, and thinking_adaptive have been removed.