Written by
Lena Hoffmann
Published
Reading time
11 min read
Topics
Research · Agents
For two years, teams squeezed quality out of a single model with longer and longer prompts. Routing each request to the best model for the job gets better results for less money.
Three signals
Our router weighs task difficulty, latency budget and cost. Most requests are simple and go to fast, small models. Hard ones get the large reasoning models.
Research · Agents
