What AI video really costs: Wan 3.0 billed more than double its listed rate
Estimating video generation as "rate per second × seconds" does not survive contact with a bill. Measured, we were charged more than twice the listed rate. This is the record.
Wan 3.0: listed at $0.05/s, billed $0.2125 for two seconds
| Item | Value |
|---|---|
| Provider's listed rate | $0.05 per second at 480p |
| Calculated from the listed rate (2s) | $0.10 |
| Actual billed amount (480p, 2s) | $0.2125 |
| Latency | Over six minutes for 480p / 2s (asynchronous) |
| Measured on | 2026-08-28 (via OpenRouter) |
The scale is different from images. At $0.019 per image for GPT Image 2, one two-second 480p clip costs about eleven images. Presenting both as "one generation" in the same UI invites accidents.
Can't I just multiply the rate by the seconds?
Two seconds should have been $0.10 and billed $0.2125. Run one clip and read the bill.
Video does not return synchronously
Images return in the same request. Video does not: you get a polling URL back and check for completion yourself. Even 480p at two seconds took over six minutes.
That matters for implementation. A press-and-wait UI borrowed from images will not hold. You need a job that is accepted, then reported on when it finishes.
Seedance 2.5: unit price is not fixed in advance
| Item | Value |
|---|---|
| Billing | Per token ($0.0000107 per video token) |
| Latency | Not measured (asynchronous) |
| Resolution | 480p / 720p |
| Duration | 4–30s |
| Frame control | first_frame / last_frame |
Seedance 2.5 bills per token, so the cost of a given clip is not known before running it. It offers something no other model here does — pinning both the first and last frame — but it does not suit work that needs a fixed cost up front.
The two compared
| Wan 3.0 | Seedance 2.5 | |
|---|---|---|
| Resolution | 480p / 720p / 1080p | 480p / 720p |
| Duration | 2–30s | 4–30s |
| Aspect ratios | 5 | 6 (includes 21:9) |
| Frame control | first_frame | first_frame / last_frame |
| Audio generation | Supported | Supported |
| Cost known in advance | Yes — from measurement, not the rate card | No |
Never put video and images in the same estimate. They are an order of magnitude apart.
How to estimate
- Do not use the published rate. Run one clip and read the bill
- Keep video and images in separate estimates — the scale differs
- Design for asynchronous completion; a synchronous UI will not work
- Hold your own per-run ceiling for token-billed models
Frequently asked
- What does Wan 3.0 cost?
- It lists at $0.05 per second at 480p, but two seconds at 480p billed $0.2125 — more than double the $0.10 the rate card implies.
- How long does video generation take?
- Over six minutes for two seconds at 480p on Wan 3.0. Generation is asynchronous and returns a polling URL.
- How much more expensive is video than images?
- Measured, one two-second 480p clip costs roughly eleven GPT Image 2 images.
