AI providers bill by the token, a small unit of text that the model reads or writes. The price of each token can fall while the total bill for AI rises.
Cheaper units invite more units.
When each request costs less, more requests are made, more text is sent with each request and wasteful repeat requests go unchecked. A falling unit price removes the pressure that used to keep consumption disciplined, and consumption can grow faster than the price falls.
The pattern is familiar from storage and bandwidth. Unlike a storage bill, though, an AI bill can rise sharply after a single change to the instructions a model is given.
Why output costs more than input
On most commercial models, output tokens are priced higher than input tokens. A change that makes responses longer can therefore cost far more than a change that makes prompts longer, and teams optimizing for prompt brevity are often working on the cheaper half of the bill.
The share of a token budget that goes to waste varies widely between organizations, so it has to be measured in each case before any figure is relied upon.
What to manage instead
The number worth putting on a dashboard is tokens per completed task, by task type. It survives price changes, it is comparable month to month, and it rises when something has gone wrong, which a total bill alone may not reveal.
A falling unit price can flatter the total bill for some time, while tokens per completed task continue to show the underlying trend.