llm 0.32rc2
Simon Willison's Weblog Simon Willison ● Covered by 14 sources
Simon Willison's LLM tool just bumped its default model from GPT-4o mini to GPT-5.6 Luna, and added a quick way to test any OpenAI-style endpoint. Small update, but it changes what happens the first time someone runs the CLI with no config.
Based on reporting by Simon Willison's Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Simon Willison pushed out llm 0.32rc2 just days after RC1, mostly to patch a dependency snag, but he snuck in two changes worth noticing. The bigger one: anyone who hasn't explicitly set a default model will now get GPT-5.6 Luna instead of GPT-4o mini. That's a meaningful swap, not a cosmetic one. Luna costs a bit more, $0.20 per million input tokens and $1.20 per million output tokens versus 4o mini's $0.15/$0.60, but it's newer and noticeably more capable. Willison left the door open for people who want to opt out, either back to 4o mini or down to the even cheaper GPT-5 nano at $0.05/$0.40, both via a one-line llm models default command.
The second addition is the one Willison seems genuinely excited about: a new llm openai endpoint command that lets you fire prompts, chats, or model listings at any OpenAI-compatible API without pre-configuring a model in LLM at all. He built it because he kept hitting a wall trying random Chat Completions-style endpoints and finding no quick CLI way to poke at them. Now you don't even need to install LLM proper; a uvx one-liner will do it, pulling in tools and everything. He demoed this against a local LM Studio instance running Gemma, asking it to check its own version and the time, tool calls included.
What stands out here is how much this update is about lowering friction rather than adding raw capability. Defaulting to a better model matters for the huge number of people who never touch config files. And the endpoint command turns LLM into a general-purpose probe for the entire OpenAI-compatible ecosystem, LM Studio, local servers, whatever else speaks that dialect, without demanding commitment upfront. Small, practical, very Willison.
My take — AI-written commentary, not fact-checked reporting
I like this update precisely because it's unglamorous. Nobody needs another benchmark chart; people need tools that quietly get less annoying, and swapping in a better default model while adding a zero-friction way to test random endpoints does exactly that. The bigger pattern here is that the real innovation in the LLM space right now isn't always a new frontier model, it's tooling that makes the fifty existing ones easier to actually use.
Read more about this at: Simon Willison's Weblog
