Tiered AI model routing groups models by expected role, such as fast routine work, general-purpose tasks and expensive high-capability reasoning. A classifier, explicit rule or user choice assigns each request to an appropriate tier.
A tiered design should include escalation when a task fails or reveals greater complexity. Evaluation must measure complete outcomes because an inexpensive first attempt followed by several retries may cost more than choosing the stronger tier initially.
ELI5
Tiered AI model routing groups models into roles such as fast routine work, general tasks, and expensive high-capability reasoning. Rules, classifiers, or user choices send each request to the most appropriate tier.
For example, a simple extraction can use a small model while a failed or difficult case escalates to a stronger one. Evaluation should count retries because several cheap attempts may cost more than choosing the stronger model first.
What tasks belong in a fast-model tier?
Well-scoped edits, searches, tests and other repeatable operations belong there when evaluation shows that the model completes them reliably.
How does a workflow escalate between tiers?
It can escalate after failed verification, detected ambiguity, explicit user choice or a classifier that identifies greater reasoning or context needs.
