Most platforms that resell AI models never tell you what those models cost them. We spent a day measuring ours, and the answer was uncomfortable enough that we are publishing it.
Most platforms that resell AI models never tell you what those models cost them. We spent a day measuring ours, and the answer was uncomfortable enough that we are publishing it.
Three of the numbers we priced against were not real. One was a fee that did not exist. One model had been selling at three times its correct price.
## Why a wrong cost is worse here than elsewhere
We are a reseller. Every model on AetherWave is bought wholesale and sold in credits, and the pricing formula is one line: your price is the provider cost times 1.4. That formula is only as honest as the cost going into it.
A wrong cost is not a rounding error. It is a wrong price on every generation, forever, and nothing breaks to tell you.
## What we believed
Our configuration made three specific claims about the Seedance family, each with a confident comment explaining its derivation.
**That a video reference triggers a fixed analysis fee.** 150 credits on any generation using one. The comment cited two figures read off the provider's playground: 250 credits for five seconds, 375 for ten, which fits "125 fixed plus 25 per second."
**That a video reference is cheaper per second.** Our tables priced it as roughly a 40 percent discount.
**That generated audio carries a surcharge.** 33 percent on most variants, with one model flagged at 4x and its audio switched off by default to avoid selling at a loss.
All three were wrong.
## How we found out
The provider exposes a field called `creditsConsumed` on its task-status endpoint. It reports exactly what one generation cost, after the fact.
It had been there all along. Nobody had looked.
Before finding it we tried measuring by reading the account balance before and after each call. That method is worse than it sounds: it assumes nothing else bills the account during the window, and on a live account that is false. One early reading came back at exactly double the true cost because another job landed inside the window. We built a confident conclusion on it and had to retract the whole thing an hour later.
The general lesson: a measurement that can only err in one direction is safe at its floor and dangerous everywhere else. The per-task field has no such problem.
## What was actually true
We ran a controlled grid. Every model, both resolutions, with and without a video reference, plus audio, at a fixed duration.
**The fixed fee does not exist.** A four-second job at 480p with a video reference costs 61.2 provider credits in total. The fee alone was supposed to be 125. There is no fixed component in any cell. Cost is linear in duration everywhere.
**Video references cost more, not less.** Between 31 and 42 percent more depending on the model. We had the sign backwards, so those generations were underpriced on the rate while carrying a 150 credit charge for something imaginary.
**Audio is free.** Across all three Seedance 2.0 variants, a generation with audio cost exactly what one without cost. The surcharge was invented, and so was the 4x figure that had kept a model's audio disabled.
Measured wholesale cost, in provider credits per second:
| | 480p | 480p + video ref | 720p | 720p + video ref |
|---|---|---|---|---|
| **Grok Imagine** | 2.40 | n/a | 4.50 | n/a |
| **Seedance 2.0 Mini** | 3.80 | 5.40 | 8.20 | 11.25 |
| **Seedance 2.0 Fast** | 11.70 | 15.30 | 24.80 | 33.75 |
| **Seedance 2.0** | 19.00 | 25.88 | 41.00 | 56.25 |
| **Seedance 2.5** | 28.00 | 33.88 | 63.00 | 76.23 |
Two prices survived measurement untouched: Seedance 2.0's base rates, and Grok at 480p. Everything else moved.
## What this did to prices
Most went down. Several went down a lot. Figures are credits for a four second clip.
| | before | after | |
|---|---|---|---|
| Mini 480p with video ref | 186 | 32 | **-83%** |
| Mini 720p | 116 | 48 | **-59%** |
| Mini 480p | 56 | 24 | **-57%** |
| Fast 480p with video ref | 202 | 88 | **-56%** |
| Fast 720p | 188 | 140 | **-26%** |
| Seedance 2.0 720p with video ref | 290 | 316 | +9% |
Seedance 2.0 Mini had been selling at between two and three times its correct price. Not because anyone decided that, but because its cost had never been established. It sat in our records as `null` and a guess filled the gap.
Two prices went up: Seedance 2.0 with a video reference, and Grok at 720p, where we had been undercharging by one credit per second on the second busiest model on the platform. We are publishing those too.
## Which model to actually use
The measured numbers change the advice, including advice we had been giving.
**Grok Imagine is the cheapest video we sell.** 4 credits per second at 480p, 7 at 720p, and it runs from 6 to 30 seconds. Its image-to-video variant accepts up to seven reference images. If you are iterating on an idea, producing social clips at volume, or working from reference stills, start here. Nothing else we carry is close on price.
**Seedance 2.0 Mini is not the budget option.** We assumed it was, twice, and it is not: Grok undercuts it at both resolutions and takes references too. What Mini genuinely offers is first-and-last-frame interpolation, plus video and audio references, cheaply. Reach for it when you need to control where a shot ends, not when you need to save money.
**Seedance 2.0 Fast and Seedance 2.0 are quality steps.** Fast turns around in one to two minutes and is the workhorse for finished social work. The full 2.0 model is better again, and its pricing was the one thing we had right from the start.
**Seedance 2.5 is for shots with requirements.** It is the most expensive per second and often still correct, because three of its capabilities cost nothing extra:
- **Image references are free.** Up to nine of them, at the base rate.
- **Generated audio is free.** Every other model in the family either charges or has it off.
- **It runs to 30 seconds**, where 2.0 stops at 15.
The reference behaviour is the real differentiator. We gave 2.5 four of our own album covers and asked for a product shot. It held all four accurately enough that the artwork is identifiable in the finished video. When we asked an image-editing model to do the equivalent, it ignored the references entirely and returned the original artwork untouched.
If a shot needs a specific product, character or piece of artwork to survive intact, 2.5 is the model that does it, and the references are free.
**One rule across the family.** Video references cost 31 to 42 percent more. Image references on 2.5 cost nothing. If you can express the idea with stills instead of a clip, do that.
## What we changed on our side
The audit tool lives in our codebase now rather than being a one-off, so any model's true cost is one command away. The pricing tables were rebuilt from the measured grid. The imaginary fee is deleted.
We also found that our model picker kept a separate copy of the rate tables, which had drifted below what the server charged. For some models the price shown before you clicked generate was 16 percent lower than the amount deducted afterward. That is fixed, and it is the change we are least comfortable having needed to make.
## Why publish this
Because the alternative is that you take our pricing on faith.
We would rather tell you that one model was priced at three times its cost, that we had a fee in our system that no provider ever charged, that we twice recommended the wrong model because we read our own configuration instead of testing it, and that the correction moved most prices down.
Anyone can publish a rate card. Fewer will publish the audit that proved the last one wrong.