Skip to main content
A cost cap. Requests ask for openai/gpt-4o as usual. Once that model has cost more than €50 in the last seven days, they run on openai/gpt-4o-mini instead, and go back to normal at the start of the next week. Nothing in your application changes. It keeps sending openai/gpt-4o.

​
The graph

Four nodes.

​
Building it

  1. Open Workflows in the console and create one. Give it a name you will recognise later, such as gpt-4o cost cap.
  2. Add a Model trigger and set its model to openai/gpt-4o. This is the name your application sends.
  3. Add a Metric check and connect the trigger to it. Add one condition: metric cost, operator greater_than, threshold 50, window 7 days. Leave inheritance on automatic, so it measures the triggering model.
  4. Add a Model selection set to openai/gpt-4o-mini. Drag from the metric check’s success output to it. Success means the condition passed, which here means you have gone over budget.
  5. Add a second Model selection set to openai/gpt-4o. Drag from the metric check’s failure output to it. This is the normal path.
  6. Apply changes, then enable the workflow from the list.

​
Confirming it works

Send a request for openai/gpt-4o and open Metrics. The recorded model is the one that actually ran, so under budget you see openai/gpt-4o and over it you see openai/gpt-4o-mini. If both show the original model, the workflow is not routing. The usual causes are that it is not enabled, that the trigger name does not exactly match what your client sends, or that a node in the graph failed validation. See the warning about unfinished nodes in the node reference.

​
Why the second selection node

You could leave the failure output unconnected and the request would run on openai/gpt-4o anyway, because a workflow that reaches no selection node changes nothing. Wiring it explicitly is still worth it. It makes the normal path visible on the canvas, and it gives you one place to change if the default model ever moves.

​
Variations

Swap the metric for error_rate over 1 hour to fail away from a model that has started returning errors. Swap it for latency_avg to shed load onto something faster under pressure. Add an Image detection node before the selection to send only requests carrying an image to a model that can see.