Two AI answers can use the same model and still cost different amounts.
A longer prompt burns more tokens, memory, and accelerator time.
Batching can make hardware more efficient, but waiting for a batch adds latency.
So the bill is not just the model. It is the shape of the request.