LLMPrice.io

Explaining an AI Bill That Moved Without a Usage Increase

A client calls because the invoice moved and nothing else did: same feature, same request volume, same month as every other month. The bill did not move because they used the system more. It moved because the meter underneath it changed, and that happens two distinct ways. A provider can cut or raise the published rate on a model you are already billing, or a provider can retire a model version and quietly route the same calls to its replacement at a different price. Either one produces the same phone call, and the worst answer is telling them you will look into it three separate times before you have anything to say. This is a normal event on a metered dependency, not a sign that something in your system is broken, and it is easiest to explain the first time it happens rather than the third.

Find out which of the two happened

Check two things before you write a word to the client. First, is the model identifier your code calls still the one you started with, or did the provider move an alias onto a newer version behind the scenes without changing the name you see in your own logs. Second, pull the rate you are actually billed at and compare it to what the same model publishes today: a straight rate change shows up as the identical identifier at a different number, with nothing else about the call any different. Rate changes are common, not rare exceptions you can assume away when a bill moves. We hold a recorded rate change for 28 first-party models in our archive, which reaches back further than the set of models we price today, and the most recent change on record is 4 days old: DeepSeek Flash's output rate fell 9.1 percent on September 14, 2026, and Mistral Medium's output rate fell 7.4 percent on September 4, 2026. None of those needed a contract change on anyone's end; the number underneath the contract changed on its own, on the provider's schedule, without asking either of you. The rate table on this site carries what each model publishes today, so the comparison is one lookup rather than a support ticket filed with the provider.

Show the rate and the usage as two separate numbers

Do not hand the client one changed total and an explanation in prose. Show what the same request volume would have cost at the old rate, next to what it actually cost at the new one, on one line each. That single comparison answers the only question they actually have, which is whether this is going to keep happening for reasons that have nothing to do with what they are asking the system to do. Our own monthly index is built the same way and for the same reason: it holds the request shape fixed across two months and reports only the rate move underneath it. On that measure an agent-shaped workload is down 48.4 percent since July 2024, on the identical request shape throughout. That is the same decomposition you are doing for one client's invoice, run at the scale of the whole market instead of one account, and it is worth showing them that the method is the same rather than something you invented to soften the news.

Decide who keeps the difference, before the next invoice, not after

A rate cut you do not pass through quietly becomes a margin increase nobody agreed to. A rate rise you absorb quietly becomes a loss you eventually have to explain all at once, in a worse conversation than the one you are avoiding now. State once, plainly, whether the new rate is being passed through or whether you are holding the quoted number, and for how long you can afford to hold it. If the contract already carries a ceiling, a rate cut needs no conversation at all: it simply widens the room under that number and nobody needs to be told about slack they are not paying for. A rate rise that threatens the ceiling does need one, and it belongs before the ceiling is breached, not inside the email explaining why it already was. Either way, tell them in the same plain terms you would use for any other metered cost: what moved, by how much, and who is paying for the difference from here. The guide on marking up API costs covers which structure makes that conversation easy, and quoting fixed-price work covers the ceiling itself.

Related