Why it mattersA team on llm now pays half as much per call on the default model. The endpoint command removes a manual config step for every new OpenAI-compatible provider.
Simon Willison has released llm 0.37. The LLM command-line tool and Python library gets around two months of changes in this release. Willison also notes a pattern that most other maintainers complain about. Twelve people contributed fixes to this release. Almost all of the fixes were written with the help of an AI coding tool.
What the release changes
The default model is now gpt-6-luna. Willison writes that this is half the price of the previous default, gpt-5.6-luna. A team that has not set its own default now pays less per call on existing scripts. The change lands with no code edit on the user's side.
The release also adds a new command. llm openai endpoint add registers any OpenAI-compatible endpoint. The command removes a manual edit to a config file for each new provider that speaks the OpenAI API shape. A team that uses Groq or Together can add it from the command line now. The same works for any local server with an OpenAI-compatible port.
Two new models are included. gpt-6.1-sol and gpt-6.1-terra are both callable with no extra setup. The embedding support was changed too. Different binary inputs can now carry different content-type values. Willison writes that this is needed for models like Gemini Embedding 2. That model accepts text, audio, images, and video in one call.
What Willison says about AI-assisted contributions
Open source maintainers often say they are getting more pull requests written with AI help. Willison says in the release note that llm has felt the pattern most of his projects. He writes that he does not find the volume to be a problem for him.
The reason he gives is a workflow. Willison says he uses Opus 5.5 or GPT-6 Astra to help with triage. The AI reads a pull request and flags the most useful fix. Willison then decides whether to land the change as written or write his own fix for the same issue. He says it felt lighter than he expected.
What the note does not cover
Willison does not publish a side-by-side comparison between gpt-6-luna and gpt-5.6-luna on any benchmark or task. The price difference is clear. The quality trade for the cheaper default is not stated. A team that cares about quality on one task should run its own test first. The default change makes a sensible test cheap.
The release is a point release in a long-running tool. It does not change the project's licence, its storage model, or its command names. A user on 0.36 can upgrade in place. A user who had set a project default is unaffected by the default change.
Source
Release note by Simon Willison.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
