Model Escalation

Simple Definition

Model escalation is moving a task from a cheaper, faster, or less capable model to a stronger model when the task exceeds a defined threshold for complexity, uncertainty, failure, or risk.

Plain-English Explanation

The typical workflow starts a task on a cheaper model, evaluates the result, and escalates to a stronger model if confidence is too low, validation fails, the task turns out to be more complex than expected, the risk level is high, or the model hits a defined retry limit without success. When escalation happens, the task and its relevant context get handed to the stronger model, ideally without losing the work already done.

Model Escalation vs. Model Routing vs. Model Fallback

These three get mixed up because they all involve choosing a model, but they answer different questions at different points. Model routing decides where a task starts, based on its type or expected difficulty. Model escalation decides when a task should move to a stronger model mid-workflow, based on how it’s actually going. Model fallback is different again: it’s a backup model that takes over when the primary model is unavailable, rate-limited, or too expensive, regardless of how well the task itself is going. Escalation is triggered by the task struggling; fallback is triggered by the model being unreachable.

Example

A support agent starts every ticket on a fast, inexpensive model. If that model’s confidence score comes back low, or its proposed resolution fails an automated check, the ticket escalates to a more capable model with the original context attached, rather than the customer ever seeing a failed first attempt.

Why It Matters

Running every task on the strongest available model wastes money on requests that never needed it. Running every task on the cheapest model risks silent failures on the harder cases that actually needed more capability. Escalation lets most work run cheap by default while still catching the subset that genuinely needs a stronger model, without requiring someone to guess correctly in advance which tasks those will be.

Model Routing, the initial decision about which model a task starts on
Model Fallback, switching to a backup model when the primary one is unavailable, not when a task is struggling
Evaluation-Gated Routing, using eval results to decide which tasks are safe to keep on a cheaper model in the first place
Effort Routing, increasing reasoning depth or review on the same model, a related but distinct lever
Task Budget, the cost estimate that an escalation path should account for