Simon Willison’s llm 0.31 ships four changes: GPT-5.5 model access, a verbosity parameter for GPT-5+ models, a configurable image detail level for vision inputs, and async registration of custom OpenAI model configurations.
Each new parameter corresponds to a feature documented in OpenAI’s API, according to Willison’s post.
New model and verbosity control
GPT-5.5 is now accessible via llm -m gpt-5.5. Alongside the model itself, the release adds a verbosity option that applies to GPT-5 and later models: -o verbosity low, -o verbosity medium, or -o verbosity high. The post links to OpenAI’s cookbook documentation describing this as a “verbosity parameter” controlling how talkative the model’s responses are. This gives callers explicit control over response length behavior rather than relying on prompt-level instructions to constrain output length.
Image detail level for vision inputs
For models that accept image attachments, 0.31 adds -o image_detail low, -o image_detail high, and -o image_detail auto. GPT-5.4 and GPT-5.5 additionally accept -o image_detail original. The post links to OpenAI’s documentation on “choosing an image detail level,” which governs the resolution at which the model processes the image. Lower detail is faster and cheaper; higher detail is more precise but consumes more tokens. The original value is available only on GPT-5.4 and GPT-5.5, according to Willison’s post.
This option lets callers trade cost against visual fidelity depending on the task.
Async registration for extra-openai-models.yaml
The fourth change is a plumbing improvement: models listed in extra-openai-models.yaml are now also registered as asynchronous. Previously, custom OpenAI-compatible models defined in this file were only registered for synchronous use. The async registration allows them to participate in workflows that use llm’s async model interface, which is relevant for callers running concurrent requests or integrating with async Python code. This change is tracked as issue #1395 in the project.
The release includes no breaking changes, deprecations, or changes to the core plugin or conversation management subsystems, according to Willison’s post. All four changes are additive.