llm-anthropic 0.26
Simon Willison Simon Willison ● Covered by 2 sources
Simon Willison's LLM tool just got an update supporting new Claude 5 models plus built-in web search and code execution.
Simon Willison has pushed out version 0.26 of llm-anthropic, the plugin that connects his command-line LLM tool to Anthropic's API, and it arrives loaded with changes that track Anthropic's own model refresh. The headline addition is support for three new models: claude-fable-5, claude-sonnet-5, and claude-opus-5, all wired in through issues #75 and #76.
The bigger structural shift is around tools. Willison has added server-side tool support for WebSearch, WebFetch, CodeExecution, and AnthropicMCP, all accessible through LLM's -T flag or the tools= parameter in Python. That replaces the older -o web_search* options entirely, so anyone with scripts built on the old syntax will need to update them. It's a cleanup that mirrors how Anthropic itself has been pushing these capabilities server-side rather than leaving developers to bolt them on.
Thinking mode also got simplified, and this is where the release notes get genuinely useful. Extended thinking is now just thinking and thinking_effort, with five levels: low, medium, high, xhigh, and max. Claude 5's Sonnet and Opus variants think by default now, though you can turn it off with -o thinking 0. Fable 5, interestingly, always thinks — no opt-out. Three older options, thinking_budget, thinking_display, and thinking_adaptive, have been removed outright, which suggests Anthropic's API itself dropped that granularity in favor of the simpler effort scale.
Under the hood, this release requires llm>=0.32, and that dependency brings streaming improvements: reasoning, tool calls, tool results, and server-side tool outputs now all stream as typed events rather than raw text blobs. For CLI users, reasoning output now prints to standard error by default, with a new -R/--hide-reasoning flag to suppress it from both the terminal and saved logs if you'd rather keep transcripts clean.
None of this is flashy, but it's the kind of maintenance work that keeps a widely used open-source tool usable the moment a model provider changes its API underneath it. Willison has built a reputation on shipping these updates fast, often within days of a new model landing.
My take
Credit where it's due: most people building on top of frontier model APIs would let their tooling rot for weeks after a provider shifts its parameters, but Willison ships a same-week patch almost every time. That responsiveness is arguably more valuable to working developers than any single new Claude model, and it's a quiet argument for why small, single-maintainer open-source tools sometimes beat bloated SDKs from the big labs themselves.
Read more about this at: Simon Willison
