If you like using Claude from the terminal, August 4th was a good day. Simon Willison shipped two releases back to back: LLM 0.32 — which he calls the most significant version since the project started — and right alongside it llm-anthropic 0.26, the plugin that brings Claude into that tool.
What LLM 0.32 brings
The headline change: LLM now shows the reasoning traces of thinking models. Run against a model that reasons, and you’ll see that thought process on standard error — visible, but cleanly separated from the actual answer. On top of that come server-side tools from various providers, redesigned content-addressable SQLite logs, and support for the new GPT-5.6 family. The cheap GPT-5.6 Luna is now even the default model for a plain llm prompt.
The Claude part: llm-anthropic 0.26
The more interesting piece for us is the Anthropic plugin. It adds the three current models — claude-fable-5, claude-sonnet-5, and claude-opus-5 — and brings server-side tools straight to the command line: WebSearch, WebFetch, CodeExecution, and AnthropicMCP, reachable through the -T flag or via tools= in Python. The old -o web_search options are gone, replaced by the tidy -T WebSearch.
The biggest everyday win, though, is the cleaned-up thinking control. Instead of the old options for budget, display, and adaptivity, there are now just two knobs: thinking and thinking_effort — with levels low, medium, high, xhigh, and max. The Claude 5 models think by default. On Sonnet 5 and Opus 5 you can turn that off with -o thinking 0, while Fable 5 always thinks. And reasoning, tool calls, and tool results now stream as typed events — something you can process cleanly instead of parsing a wall of text.
My take
This isn’t a fireworks show of new features — it’s solid tool maintenance, and that’s often the more valuable kind. If you use Claude scriptably in pipelines, automations, or small tools, you get two things that matter here: visible thinking that doesn’t pollute your output, and server-side tools you previously had to cobble together yourself. The fact that thinking control shrank from a dozen options down to two says a lot about how mature these models have become. Fewer buttons, more clarity. An llm install -U llm-anthropic is worth it.
Sources: