Skip to main content
Spending caps stop the agent from starting an expensive run without your confirmation. They are configured in Settings → Spending caps.

​
How the cap works

A run group asks for confirmation when both conditions are true:
  1. The projected cost of the turn, including this run, is above your safety cap.
  2. The run’s media type has confirmation turned on.
Runs below the cap, and runs of a media type with confirmation off, start automatically. Runs with an unknown cost never ask.

​
Defaults

​
Cost estimates

The agent estimates cost from the model’s price and from historical cost per request for that endpoint. The estimate is conservative. It uses a high percentile of past requests, so real costs are usually lower than the estimate. The agent also has a pricing tool and must check a model’s real price before it quotes a cost to you.

​
When a run is waiting

An approval alert appears in the chat with the estimated cost. Until you approve or reject it, the agent cannot submit further runs in that chat. A blocked submit shows the message “Blocked because a previous generation is awaiting cost approval.” In a model comparison, waiting columns show Awaiting approval.

​
The spend chip

Next to the composer, a chip shows the number of generations and the total cost of the chat. Hover it for a breakdown By model or By modality. A dot on the chip means some runs are not priced yet.