A new routing method asks fewer questions before it starts saving you money.
LLM routers send each query to the cheapest model that can still handle it well, but training one usually means running every candidate model on a stack of historical queries first, and that data collection isn't free. Researchers behind SaveRouter noticed routing accuracy tends to plateau long before all that feedback gets collected, meaning most teams are paying for supervision they don't need. Their framework instead selectively picks which query-model pairs are worth testing and shares what it learns across similar queries, only drilling into query-by-query detail when it matters. Across four routing benchmarks, SaveRouter matched or beat existing routers using only about a third to two-fifths of the usual training feedback, cutting the query volume needed to break even by up to roughly 9.5 times versus the fastest conventional approach.
The interesting part isn't the accuracy, it's the accounting. Most routing research measures how well a router picks models, not whether the setup cost was worth it, and that's the number that actually determines if routing saves anyone money. The paper also found that the supervision level that produces the best long-run routing isn't necessarily the one that pays back fastest, which is a genuinely useful distinction for anyone deciding how much upfront testing to budget.
It's a narrow, unglamorous fix, but it targets the part of the routing pitch that vendors tend to skip: the bill you rack up before you save anything.