One price for 'a search' hides the only decision that matters: how many pages to read, and how carefully.
A web search for an AI agent is not one operation. It is a chain: discover candidate pages, fetch them, render the ones that only exist after JavaScript, extract the text that matters, rank, and write. Every link in that chain can be done quickly or thoroughly, and the difference is visible in the answer.
So a single price for a search is an average that hides the decision. Easy questions subsidise hard ones, and the developer gets no say.
Four depths
Gistly exposes the dial directly: quick, standard, deep and max. Each reads more sources, renders more of them in a browser, and allows more time before the deadline. Each has a fixed price, published on one page, and the price of the call you made comes back in the response.
What stays fixed
The shape of the answer does not change with depth: one written answer, the sources it was written from, and a flag when the source list was cut short. The agent can start cheap and go deeper only when the first answer cites too little. That is a decision the agent can make; a single price takes it away.
Where the deadline lives
Every depth carries its own deadline, and the documentation says what an HTTP client's timeout should be above it. A search that cannot finish is refused with a clear status and is not charged. Predictable failure is part of a fair price.