Mitte showed me 150 credits. It charged 1,120.

A field note on a credit meter where the displayed price is sometimes the total and sometimes a per-second rate, with nothing on screen telling you which.

The generate button said this avatar clip would cost 150 credits. The usage log took 1,120. Same clip, nothing changed between the quote and the charge.

I checked it three ways before I trusted it. The log showed minus 1,120. My balance moved from 8,673 to 7,553, which is 1,120. The lifetime counter went up by 1,120. All three agreed. A shown 150 became a billed 1,120, a 7.5 times jump on one clip.

Veed Fabric, 720p. Shown: 150 credits. Charged: 1,120. Output 1312×736, no watermark.

My first thought was a bug. It was not. Here is the part that took a while to see.

I had run the same avatar at 480p first. That one showed 80 credits and charged 720. So the cheaper setting also ran far over its quote, by a similar multiple. Two clips, two quotes, both off by seven to nine times, in the same direction.

When numbers miss by the same shape twice, it is usually not an error. It is a unit.

Line the clips up against their length. The 480p clip ran a little over eight seconds on a shown 80, and charged 720. The 720p clip ran just under seven seconds on a shown 150, and charged 1,120. In both, the charge sits close to the shown number multiplied by the clip’s length in seconds, within about ten percent. The number on the button was not the price of the clip. It behaves like the price of one second. Talk for seven seconds and you pay for seven of them.

I have not confirmed this with Mitte, and I am holding it as a read of the math, not a statement about their billing. But it fits beyond the avatars. The image tool in the same review showed 45 credits for a six second clip and charged 280, and 45 across six seconds is 270.

Here is where it stops being a clean theory and becomes the actual problem.

That per-second read cannot explain the generations that billed exactly what they showed. GPT Image 2 quoted 20, 110, and 410 across its three quality tiers, and charged 20, 110, and 410. Veo quoted 200 and took 200. Seedance quoted 968 and took 968. On those, the displayed number was the whole price. On the avatars, it was a rate. Same product, same screen. The same kind of number meant two different things, and nothing on the button told you which.

GPT Image 2, same prompt, 4K. Shown and charged, identical: 20, 110, 410 credits. Here the number was the total.

So the honest version of the finding is not that the meter lies. It is quieter and harder to plan around. The displayed credit number does not carry a fixed unit. You cannot look at a quote and know whether it is the bill or a fraction of it. The only way to find out is to run one and read the log after.

For the record, the output itself was straight. The 720p clip came back at 1312 by 736, a real 720p at 25 frames a second, with no watermark. The resolution was honest. It was the price tag that was ambiguous.

The first principle here is dull and load bearing. In a metered tool, treat every displayed price as unverified until you have run that exact generation once and checked the charge against the log. Do it per model and per setting, because the unit can change between them. A quote is a hypothesis. The log is the receipt.

I took Mitte’s full pricing and data terms apart in the FSR review, including how the displayed numbers line up against the charges. If you want the teardown, it is here: [Mitte review on FSR]

The field note version is one line. The number on the button is not a price until the log agrees with it.