>or perhaps to route certain requests to Haiku or Sonnet instead of using Opus for everything, to cut down on the compute

You can point Claude Code at a local inference server (e.g. llama.cpp, vLLM) and see which model names it sends each request to. It's not hard to do a MITM against it either. Claude Code does send some requests to Haiku, but not the ones you're making with whatever model you have it set to - these are tool result processing requests, conversation summary / title generation requests, etc - low complexity background stuff.

Now, Anthropic could simply take requests to their Opus model and internally route them to Sonnet on the server side, but then it wouldn't really matter which harness was used or what the client requests anyway, as this would be happening server-side.