llm 0.38
- Source
- github.com
- Author
- simonw
- Date

Context
Version 0.38 of , Simon Willison's command-line tool and Python library for working with language models, adds OpenAI's new GPT-6.2 Astra and GPT-6.2 Sol as gpt-6.2-astra and gpt-6.2-sol (release notes). The rest of the release is bug fixes and a documentation change, and two of them change what users can rely on.
The first affects programs that give models tools to call. LLM describes each tool to the model with a schema, a structured list of its inputs, and strips out the title labels that the Pydantic library adds automatically. That cleanup also removed title keys that belonged to real values, such as a default value of {"title": "A document"}, so the model was told about a default that was missing its key. Tool schemas now keep those values intact. Separately, setting LLM_OPENAI_SHOW_RESPONSES=1, which logs raw requests and responses, previously made every async OpenAI prompt fail before a request was sent; async calls now work and are logged like synchronous ones.
The second is a documentation warning that matters for anyone with plugins. Upgrading LLM with uv tool upgrade llm, pipx upgrade llm or brew upgrade llm removes plugins installed through llm install, and the docs now recommend llm install -U llm to upgrade while keeping them. Other fixes make commands that never call OpenAI, such as llm --version, start faster by loading the openai package only when needed, and read templates and fragment files as UTF-8 so non-ASCII characters no longer cause errors on systems such as Windows.
- Upgrading llm with uv tool upgrade, pipx upgrade or brew upgrade removes plugins installed through llm install, so the docs now recommend running llm install -U llm to upgrade while keeping them.
- Templates, -f fragment files and --functions Python files are now read as UTF-8 first, falling back to the system encoding, which fixes UnicodeDecodeError crashes on systems such as Windows when files contain non-ASCII characters.
- A Toolbox subclass that called super().__init__() on a parent with its own constructor used to have its arguments overwritten by the parent's, so instances differing only in a subclass argument were logged as the same toolbox; 0.38 records the subclass arguments.
- The async logging bug was total, not intermittent: with LLM_OPENAI_SHOW_RESPONSES=1, every async OpenAI prompt failed with an invalid http_client error before any request went out, and async traffic is now logged the same way as sync.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
Checking sign-in…
Loading comments…


