Disclosure: I'm the author of Lynkr, the open-source proxy whose internals this post walks through. All code shown is real and Apache-2.0 — read it here.
The most expensive default in AI coding tools is that model choice is a setting, not a decision. You pick a model once; every request — "what does git stash do?" and "refactor this auth module" alike — goes there. Routing each request to the cheapest model that can actually handle it is worth 50%+ of most bills, but it only works if the "can actually handle it" judgment is reliable. Get it wrong downward and a small model fumbles your file edits; get it wrong upward and you've saved nothing.
Here's how Lynkr makes that judgment, in enough detail that you could reimplement it.
Why not just count tokens?
The obvious heuristics fail in both directions:
