It cost more than I expected

Two things explain most of the surprises, and they are worth checking before anything else: the run was not on the tier you thought, and a spend limit stops a run after the step that crosses it, not before. Work down the list until one matches.

1. Check which tier the run was on

The receipt names it. If it says Max and you meant Fast, that is your answer — tiers differ in what they cost per unit of work, so the same job is a different total on each.

The trap is that the engine picker on the home screen is your account default, not a one-off choice. Change it there for one cheap errand and every new chat afterwards runs on it, with no reminder. The picker inside an open chat is scoped to that chat, and Retry with a different engine also changes that chat’s tier from then on. There is no “just this message” scope. Switching models lays out all four places and which scope each one has.

2. Count the turns, not the runs

The spend ceiling is per turn, not per conversation. A conversation of six turns can spend several times it, because each turn starts the count again. A long back-and-forth is the most common way to a total that feels out of proportion to any single answer in it.

3. Expect the limit to be passed, not held

In a chat the number you type above the composer is enforced — but the check happens after a step has already been paid for, so the turn stops on the step that takes it past your number and the total is your limit plus that step. A total a little above what you set is the mechanism working, not a fault. A blank field really is “no limit”: nothing caps that turn’s spend, and the line under the composer says no spend limit until you type a number. Work you hand off does not carry that number at all — a task’s own limit is a running total, checked between steps against money already spent, so it cannot prevent an overrun and never fires on a job that finishes in one push. Spending limits is the full account, and nothing on this page softens it.

4. Look at what the work actually involved

Long work gets expensive faster than the length suggests: every step carries the conversation so far, so step twenty is reading much more than step two. Large attachments, many files, and repeated searching all compound the same way. Why it cost that much breaks a real total into its parts and is the page to read if the first three checks did not explain it.

5. Check the receipt line by line

The receipt is the record — it is accurate about what you were charged, whatever any limit did or did not do. It shows each priced piece of work rather than a single number, which is usually enough to see which part of the job was the expensive one. Reading a receipt explains every field and where receipts appear.

6. Compare against the actual prices

If you want to check the arithmetic rather than the behaviour, the per-model prices and the model each tier runs by default are on Model prices. Quoting a number here without the tier beside it would be misleading, so that page is the only place we print them.

What to change next time

• Set the tier deliberately, in the chat rather than on the home screen — that keeps the choice from following you into every future conversation.
• Hand long work off as a task. Not for the per-turn limit — that number does not travel with it — but because a task has a running total displayed against its own limit as it goes, and it is built for work longer than a single turn.
• Ask for one thing per run. A run that has to figure out what you wanted spends money figuring it out — see Writing a good request.
• Watch the running total, not the limit. The limit is what you asked for; the total is what happened.

Else will never ask you to raise its own spend limit — a request to do so is rejected before it can reach you. If you see one, it did not come from us.