# Reserve, then settle: pricing a call before you know its cost

> An AI turn's cost is unknown until it ends; a SERP's charge arrives with the response. Reserving a ceiling and settling the real charge prices both up front.

Source: https://netweft.com/blog/reserve-then-settle
Updated: 2026-09-12

**netweft reserves credits at the rate card's price before a paid call goes out, then reads the real charge off the response and settles against it. The reservation is a ceiling, so a cheaper operation releases the difference, and a workspace that cannot cover it is refused before anything is bought.**

Two operations in netweft have a cost that is genuinely not knowable in
advance. A SERP lookup's real charge arrives with the vendor's response. An
assistant turn's cost depends on how many tokens the model decides to produce.

Both still have to be priced *before* they run, because the product's whole
claim is that nothing spends without quoting. The mechanism that resolves that
is reserve-then-settle, and it is worth writing down because the obvious
alternatives are all worse.

## What it does

1. The rate card names a price for the operation. That price is a **ceiling**,
   chosen to be at or above the realistic worst case.
2. Before the call goes out, that many credits are reserved against the
   workspace. A workspace that cannot cover it is refused here — with
   `available` and `required` in the body — and nothing is bought.
3. The call runs.
4. The real cost is read off what came back and settled against the
   reservation. If it cost less, the difference is released.

The consequence is the one that matters to a person using it: **you are never
charged more than the number you were shown, and often less.**

## Why not bill afterwards

Because then the refusal has nowhere to happen. A product that bills in arrears
discovers that a workspace could not afford the call at the moment it sends the
invoice, which is the moment it is too late for everybody. There is no version
of "we already spent your money, please pay us" that is good for either party.

Reserving first moves the failure to the only point where it is still cheap: in
front of the button, before the vendor is reached.

## Why not quote the exact price

Because for an assistant turn there isn't one. A ceiling that is occasionally
generous is a ceiling that can be quoted honestly; an "exact" price that is
really an average is a number that is wrong in both directions and admits to
neither.

The [assistant's per-turn charge](/pricing) is the clearest case. It is stated
as a ceiling, on the rate card, in those words — and the settlement afterwards
is what makes the ceiling fair rather than a rounding-up.

## What it costs us

Reservations hold credits that the operation may not use. A workspace running
close to its balance will occasionally be refused an operation it could in fact
have afforded, because the ceiling was conservative and the real cost would
have fit.

That is a real cost and it is the right way round. The alternative failure —
letting a call through that turns out to be unaffordable — is one where the
money is already gone.