Simon Willison's llm 0.31 brings GPT-5.5 into the boring test loop

The Simon Willison llm 0.31 release notes deliver a very specific kind of good news: GPT-5.5 is now addressable from the command line as simply `llm -m gpt-5.5`. It also adds a text verbosity option for GPT-5+ OpenAI models, image-detail options for attachments, and async model registration. That is not confetti; that is test harness material. A web chat is fine for vibes, but a terminal command is better when you need repeatability. Builders need to run the same prompt against multiple models, pipe files in, diff results, script batches, and hand a failing case to someone else without saying "wait, let me find the tab."

The tool’s value here is not that it proves GPT-5.5 is universally better, but that a new model can enter a repeatable developer loop quickly. Verbosity sounds like a cosmetic feature until you are paying for output tokens or trying to keep an agent from writing a novella in response to a simple renaming task. Image-detail options matter because multimodal inputs alter both quality and cost. The boring CLI flags are where evaluation discipline actually starts. A setting captured in a command can be stored in a script, reviewed in a pull request, and compared against historical output. Pick ten tasks your current model handles poorly and ten it handles well, then run GPT-5.5 across both sets to see if the upgrade helps or if it quietly regresses boring reliability. The llm tool lowers the friction between "new model exists" and "we know whether it actually helps our work."