Each vendor invoices separately. The app still has one budget.
Several vendors
One ceiling across every model vendor you call.
OpenAI, Anthropic, Google, Mistral, xAI, or a gateway you add. The model name picks the key. The project ceiling is in dollars.
What goes wrong without a ceiling.
The provider bill arrives after the traffic. Fuse is the stop in front of that call.
A model you did not turn on is refused before it is forwarded.
The price row for that model is what turns tokens into the reservation.
Three ways to keep several providers honest.
A key per provider
Store each vendor's key on the project. Fuse sends the one that matches the model.
gpt-4.1-mini → OpenAI · claude → Anthropic
Allow only what you priced
If Fuse cannot price the model, the request is 400 and never leaves.
14:02 · gemini-2.5-flash · $0.002 · 200
Same 402 for every vendor
The ceiling does not care which API would have been called. Over the line is over the line.
402 · fuse_blown · provider not called