OpenRouter has added Classifiers, a workspace feature that tags each model generation with custom labels such as department, task type, or agent complexity. The goal is to make agent activity easier to audit when one application may be calling many models for many different reasons.
Teams define a taxonomy, then use a small model to assign labels to generations. Those labels can be used to filter logs and group Activity analytics, which gives operators a clearer view of what an agent is doing and where money is being spent.
The update is aimed at a practical problem in production AI systems: raw token and model logs often show cost, but not business purpose. A support workflow, a research agent, and an internal coding assistant can look similar in infrastructure data unless teams add their own context.
Classifiers do not by themselves prove whether an output was useful or safe. They give teams a structured layer for measurement, so unusual workloads, expensive task categories, or changes in agent behavior can be spotted more quickly.