One shape
Your code speaks POST /v1/chat/completions and that is the whole
contract. The request you send is forwarded byte-for-byte — the router never
re-encodes your body, so what the provider receives is exactly what you wrote.
Every model you already pay for
An alias is one name for one model, served by an ordered ladder of providers.
chat-fast is minicpm-v today and something else
tomorrow, without your code changing.
Costed in dollars
Every request is written to a ledger with its real price — input, output,
images, video seconds — so a run is a number you can act on rather than a
usage total you have to interpret.
A budget that stops, not warns
A per-key cap is enforced before the provider is called. A request that
would overrun is refused up front, not billed and noticed later.