Model Routing
Selecting the right model for every request.
Model Routing technology encodes the policies and learned routers that send each task to the model that maximizes quality-adjusted cost. It is the scientific core behind AI Orchestrator and is informed by our empirical routing studies across open and proprietary families.
Heuristics
Fast rules for obvious task classes.
Learned Routers
Classifiers trained on production outcomes.
Cascades
Escalate from small to large models on uncertainty.
Economics
Quality-adjusted latency and cost metrics.
Routing is an ML problem
Static provider lists waste money. We treat routing as a decision under uncertainty: estimate task difficulty, predict model success, and choose under latency and budget constraints. Cascades stop early when a small model is confident.
Grounded in measurement
Our technical blog on model routing documents trade-offs across chat, extraction, code, and vision. Those recommendations ship as default policies customers can override.
Safe experimentation
Online learning is constrained by shadow traffic and human eval samples so routers improve without silently degrading critical workflows.