Two answers from the same AI can cost very different amounts to produce.
That's because each reply is billed in tokens, on top of the accelerator time, memory bandwidth, and server capacity behind it.
A 500-word answer runs more computation than a 5-word one, and a bigger model pushes more data through the chips for every token.
Batching many requests uses the hardware better, but waiting to fill a batch adds delay.
So the cost never comes from the model's name alone; token count, model size, and how busy the servers are settle it together.