Willison's llm 0.33 splits model config from prompts

Simon Willison shipped llm 0.33 on August 22, splitting saved model configuration from prompt templates so the two can be composed separately. Anyone who's copy-pasted the same reasoning-effort flag into every prompt file gets a way to stop. So does keeping near-duplicate templates that differ only by model choice.

The mechanics: llm -m gpt-5.6-luna -o reasoning_effort high --save lhigh saves a named template holding just the model and its options. Running llm -t lhigh -t pelican later stacks that config template with a separate prompt template called pelican. Model settings and prompt text stay two independently versionable pieces instead of one bundled file.

The release also closes out a rockier week for the OpenAI Python migration underneath llm. An emergency 0.32.1 patch went out a day earlier, on August 21, to fix breakage from that same migration. Version 0.33 finishes the move and switches the library from httpx to httpx2.

Two smaller changes ship alongside it. llm embed and embed-multi now accept a --key argument resolved per call instead of mutating shared model state. That means a script embedding with different keys won't leave one call's key stuck for the next. Responses API reasoning models also get a reasoning_summary option, with auto, concise, and detailed settings. The setting controls how much of the model's reasoning trace comes back in the response.