llm 0.31 makes GPT-5.5 easier to test from the terminal

Simon Willison has shipped llm 0.31, and the release is squarely aimed at builders who want to poke new OpenAI models from the command line without rebuilding their tooling first.

The headline is GPT-5.5 support: `llm -m gpt-5.5`. The more practical story is the set of small controls around it, including OpenAI verbosity settings, image-detail options for attachments, and asynchronous registration for models listed in extra-openai-models.yaml.

That is not glamorous release-note confetti. It is the stuff that determines whether a model is easy to compare, script, and drop into a real workflow instead of remaining something you admire in a launch post.

Source credit: Simon Willison’s release notes.

The workflow win is faster model comparison

For developers, llm has always been useful because it makes model testing feel like normal terminal work: quick prompts, files, scripts, aliases, and repeatable commands. Version 0.31 keeps pushing in that direction.

GPT-5.5 availability matters, but so do the knobs. Verbosity control gives builders a faster way to test whether a model should answer tersely or explain itself. Image-detail control makes multimodal runs less mysterious and potentially less wasteful. Async registration matters when models are used inside longer-running jobs instead of one-off prompts.

• try GPT-5.5 with `llm -m gpt-5.5`
• use `-o verbosity low|medium|high` to tune response length
• use `-o image_detail low|high|auto|original` where supported for image inputs
• check async behavior if your scripts rely on extra OpenAI model definitions

Who should care

The first audience is builders already running model bake-offs in terminals, CI scripts, notebooks, or small internal tools. If you are deciding whether GPT-5.5 is meaningfully better than your current default, this release lowers the friction of finding out.

It also matters for teams trying to keep model evaluation honest. A web chat test is fine for vibes. A repeatable command is better when you need to rerun prompts, compare output styles, or hand a test case to another developer without a small ceremony around browser tabs.

What changes in practice

Do not treat llm 0.31 as a reason to swap models everywhere by lunchtime. Treat it as a cleaner way to test. Run the same prompts across GPT-5.5 and your current baseline. Try low verbosity on production-shaped tasks. Check whether image-detail settings change quality or cost for your actual files.

The useful machine here is not just GPT-5.5. It is the feedback loop around GPT-5.5. The teams that learn fastest will be the ones with boring, repeatable tests ready when the shiny model arrives. Annoyingly, boring wins again.