RouteLLM vs decision models: routing between models vs deciding directly

What RouteLLM does

RouteLLM (LMSYS) trains routers on preference data to send each query to a strong model only when needed, reporting up to 85% cost reduction while keeping ~95% of GPT-4 quality on common benchmarks. It is model-selection routing: every request still runs a generative LLM, just a cheaper one when possible.

The other half of the cost story

A large share of "which LLM should answer this?" workloads are not generation at all — classify, gate, grade, match, route. For those, the cheapest path is not a cheaper LLM, it is no generation: decision models return the typed verdict directly, with output tokens free.

In practice the two layers stack: RouteLLM-style model selection for generative calls, decision models for the judgment calls around them (should this query be answered at all, which handler, what severity).

One endpoint for the judgment layer

EasyJev hosts the full System One catalog — Jev, pplx-decider, d1, Drex, Clef and more — behind one API compatible with the official typesafe-sdk, with per-request model choice and token-precise billing.