llm 0.35 adds the gpt-6-astra model to Simon Willison's CLI

On 7 September, llm 0.35 was published, the new version of the command line tool Simon Willison has been using since 2023 to talk to language models from the terminal. The release notes fit in a single line: a new OpenAI model, `gpt-6-astra`, corresponding to GPT-6 Astra.

A changelog with one bullet point is not usually news. What matters here is not the content of the entry but what it reveals about the tool's design: adding a freshly announced model cost so little that nothing else was needed to justify the release.

What llm is and why it updates like this

llm works both as a CLI and as a Python library. A call such as `llm 'summarise this text'` returns the answer on standard output, chainable with pipes like any other Unix command, and every prompt and response is stored in a local SQLite database you can later query with `llm logs`. Providers are added through separately installable plugins, and a good share of commercial APIs mirror the OpenAI format, so registering a new model often comes down to declaring its identifier.

That is why a whole version can be justified by a single note. The alternative, waiting weeks for the provider to ship its own client or for a large SDK to add the model, is exactly what this design avoids.

What the source says about gpt-6-astra and what it does not

Let us be precise. The release provides the `gpt-6-astra` identifier and the link to the OpenAI announcement. It includes no context window, no pricing and no evaluation results. Anyone who needs those figures has to go to the OpenAI page linked from the changelog itself. The only thing verifiable from llm is the string you pass in `llm -m gpt-6-astra`. Filling in the rest with guesses would be inventing, and we are not going to do that here.

Who this is useful for

Three specific profiles:

1. Teams comparing models before settling on one for production. With the `llm-anthropic` plugin installed, the Claude models are available in the same CLI, so contrasting two providers on the same prompt means changing the value of `-m` and reading the two logs.
2. People automating with shell. llm fits into cron, into Makefiles and into any script that already uses pipes, without standing up a service or writing client code.
3. People who need traceability. The SQLite log leaves an auditable record of what was asked, with which model and when, something a browser chat loses the moment you close the tab.

Where this fits if your day to day is Claude

llm and Claude Code solve different problems and should not be confused. Claude Code is Anthropic's official CLI, with skills, subagents, MCP servers, hooks and plugins, built to work inside a repository and with project context. llm is a small, multi provider piece aimed at scripting and quick evaluation. In our workflow they coexist without friction: Claude Code for the work on code, llm for firing the same prompt at four models and seeing which answers best before deciding anything.

There is also a parallel worth noting. The fact that adding a model costs one line is a consequence of interfaces having stabilised. It is the same principle behind MCP: once the contract between layers is defined, integration stops being a project and becomes configuration. The ecosystems that push that idea furthest are the ones that absorb new releases without rewriting themselves.

Our reading is a sober one. llm 0.35 changes nothing in practice for anyone not using OpenAI, but it is a good reminder that the tool that survives model churn is not the most complete one, it is the one with its extension point in the right place. It is worth having installed even if your default model is something else.