What's this about?
Coding agents run for dozens of model calls, tool uses and edits. Along the way you face a cost/quality trade-off: escalate to a stronger model when the cheap one struggles, or downshift once the hard reasoning is done. This paper measures what each switch costs, because the receiving model has to continue a trajectory it did not produce. 12 upvotes on HuggingFace.
What is a "trajectory"?
The full record of a run: every step taken, every tool result, every dead end backtracked out of. When you swap models mid-run, the new model inherits that record as text but not as understanding — it has to reconstruct why the previous model made each choice, from evidence that was never written for an outsider. The paper's framing is that this reconstruction has a measurable price, hence "tax".
Why does this matter to you?
- It quantifies something people do constantly on instinct — bumping to a bigger model when stuck — and shows it is not free.
- It is the direct counter-consideration to the routing gateway in today's Show HN, and the two should be read together.
- It suggests the right place to switch models is at task boundaries, which is exactly what that gateway's author concluded independently.