Hosted Wan 3.0 per clip vs running Wan 2.2 on a rented GPU
Sume's wan-3.0 bills $0.0625, $0.125 or $0.25 per second at 480p, 720p, 1080p. Here is the break-even arithmetic against your own GPU hourly rate.
Sume's wan-3.0 catalog row lists provider rates of $0.05, $0.10 and $0.20 per second at 480p, 720p and 1080p (dated 2026-08-24, subject to change), and Sume bills list times 1.25. A 5-second 720p clip is therefore $0.625. Running Wan 2.2 yourself costs your GPU hourly rate times minutes per clip, divided by 60. Which is cheaper depends on a number only you have: how many clips an hour your GPU really produces.
What does a hosted Wan 3.0 clip cost?
The repo's catalog constraint gives the list rates, and the docs state that every Video Router model bills list times 1.25. Multiply and you get the billed rate per second and for a 5-second clip.
| Resolution | List per second | Billed per second | Billed 5-second clip |
|---|---|---|---|
| 480p | $0.05 | $0.0625 | $0.3125 |
| 720p | $0.10 | $0.125 | $0.625 |
| 1080p | $0.20 | $0.25 | $1.25 |
How do I price a self-hosted Wan 2.2 clip?
Cost per clip is the hourly GPU rate multiplied by minutes per clip, divided by 60. The rows below use example inputs, not quotes or measurements; replace them with your own rate and your own measured minutes. Wan 2.2 TI2V-5B makes a 5-second, 24 fps 720P clip and lists a 24GB card as the minimum.
| GPU $/hour | 2 min per clip | 5 min per clip | 10 min per clip |
|---|---|---|---|
| $1 | $0.033 | $0.083 | $0.167 |
| $2 | $0.067 | $0.167 | $0.333 |
| $4 | $0.133 | $0.333 | $0.667 |
Is raw compute really cheaper?
On paper, often yes, for 5-second drafts you can keep a GPU busy on. But the comparison is not like for like. The hosted route is a newer Wan generation that accepts 2 to 30 seconds and up to 1080p, honors reference audio and video, and generates audio. The open 2.2 checkpoint is the one you can download. Your table also leaves out idle hours, setup time, failed runs, storage and the person maintaining the box.
When does each side win?
Self-hosting wins with steady, high-volume 5-second drafts on hardware you already own or can keep busy, and when the weights must stay private. Hosted wins for spiky demand, clips longer than 5 seconds, 1080p, and anyone who values not operating GPUs. A sensible test is a single 480p draft on the hosted id at $0.3125 for 5 seconds, next to one local run, before you size a fleet.
Hosted jobs also reserve funds before they run; read the Jobs and results page for how jobs settle. Rates can change, so read pricing_skus from GET /v1/videos/models rather than hard-coding the numbers above.
Sources
Related posts
More in Pricing
- WaveSpeed starting price: lowest resolution, and a 10 s clip cost
WaveSpeed's page says starting prices are list prices at the lowest resolution or quality. At those rates a 10-second clip runs $0.30 to $1.80. Table inside.
- WellSaid Starter 20 minutes vs Sume TTS per character: cost per hour
WellSaid sells narration as download minutes per plan. Sume bills per character. Cost per minute for each WellSaid plan against Sume at the same speaking rate.
- What is one AI credit worth in dollars? Runway, Vidu, Pika, Topaz
One credit is $0.01 on Runway's API, $0.005 on Vidu, about $0.011 on a Pika plan and $0.08 to $0.12 on Topaz. Convert to dollars before you compare.
- Why Sume usage rows add up to more than you spent
Summing rows from GET /v1/usage overstates spend: refunded rows keep their hold amount. Read summary.debited_usd for a thread, run or job instead.
Written by Sume