September 22, 2026
AI-agent spending can be broken down by task: Classifiers offers 8 dimensions
On September 22, OpenRouter introduced Classifiers with 8 dimensions for labeling model requests. The classifier runs after the response and does not delay the original call. Tags appear in Logs and Activity Explorer.
OpenRouter
@openrouter
TIP: automatically classify LLM requests through Jev from @typesafeai and Classifiers: https://openrouter.ai/workspaces/default/classifiers Then analyze the results in Explore: https://openrouter.ai/activity/explore
· 13.1K views
Now Activity Explorer groups traffic by task type and agent complexity, showing models and costs. The ready-made Engineering work profile marks Feature development, Bug fixing, Documentation, Refactoring, and Code review, while a custom classifier supports up to 8 dimensions.
A workspace administrator creates the classifier: Create classifier, then choose a preset or custom option, set the sampling rate, and Save and activate. Jev truncates each turn to 5 000 characters. If the model context is materially smaller, the classifier will finish without tags.
Pricing comes from the workspace. Classifier tokens are charged as regular generation against the administrator's credits, not the API key. Jev 1.13 launched on September 18, supports 32K tokens, and costs $0,042 per 1 million input tokens. The unified API requires an OpenRouter key, the base URL `https://openrouter.ai/api/v1`, and a model such as `~typesafe/jev-latest`.
On a timeout, model error, or invalid response, the API call remains unchanged, but no tag will appear in Activity.
