Why is AI video generation so expensive?
Because a video model has to generate every pixel of every frame, and the compute grows with width × height × frame rate × seconds: a 5-second 720p clip is over a hundred frames, each about as big as an image. That is why, on the same platform (Atlas Cloud), our account was charged $0.086 for four Midjourney images but quoted $0.29 for a single 5-second Seedance 2.0 Fast clip.
Where the money goes
BytePlus, ByteDance's cloud, publishes the formula it bills Seedance by: estimated tokens equal (input video seconds + output seconds) × width × height × frame rate ÷ 1024[1]. Tokens are a proxy for the compute the model spends. Every factor in that formula multiplies the bill:
- Frames. At 24 frames a second, 5 seconds is 120 frames that must stay consistent with each other.
- Pixels per frame. A 1080p frame has 2.25 times the pixels of a 720p frame.
- Seconds. Double the length, double the tokens.
- Model size. The bigger model costs more per token: BytePlus lists Seedance 2.5 at 10.70 USD per million tokens at 720p against 5.6 for 2.0 Fast[1].
- Input video. Reference clips are counted in the formula too[1].
Google's own Gemini help says the same thing in plain words: usage limits there are compute-based, and video is the most usage-intensive feature[2].
Why does a 10-second video cost so much?
Run the formula. A 10-second 1280×720 clip at 24 frames a second is 1280 × 720 × 24 × 10 ÷ 1024, or 216,000 tokens. At 1920×1080 it is 486,000. Multiply by the model's price per token and you get the bill. In Atlas Cloud's quotes for our account, 10 seconds of Seedance 2.0 Fast at 720p was $0.58, and 10 seconds of Seedance 2.5 at 1080p was $5.94: same length, about ten times the price, because of the larger model and the larger frame.
| Factor | Example |
|---|---|
| Top model tier | Veo 3.1 is $0.40 a second at 720p, Veo 3.1 Lite $0.05[3] |
| Higher resolution | Seedance 2.5 on fal.ai: $0.4730 a second at 720p, $1.164 at 1080p[4] |
| Video references | Seedance 2.5 on Replicate: $0.2312 a second without a video input, $0.9676 with one[5] |
| Generated audio | Veo 3.1 on Google Cloud: $0.40 with audio, $0.20 video-only[6] |
| Length | Our 30 s Seedance 2.5 quotes: $9.03 at 720p, $17.77 at 1080p |
Why apps cost more than the model's API price
An app that sells credits pays the model price and then adds its own costs: the language model that rewrites your prompt, storage, payment fees, renders that fail and get refunded, free credits for new users, and a margin. Credits that expire unused also raise your real cost per clip. That is why a subscription can work out at several times an API's per-second rate; Runway Standard, for example, is about $0.29 a second for Gen-4.5 if you use every credit[7]. Our cost-per-second table lays apps and APIs side by side, and how AI video credits work shows the conversion.
Will it get cheaper?
Older and smaller models already are. BytePlus lists Seedance 1.0 Pro Fast at 1 USD per million tokens against 7.0 for Seedance 2.0[1], and Google sells Veo 3.1 Lite at an eighth of Veo 3.1's 720p price[3]. Open-weight models such as Wan2.2 cost nothing to license, only the GPU to run them[8]. The newest flagship models, at the highest resolution, are what stays expensive. So the practical rule is: draft cheap, finish expensive. Make your test takes on a Fast or Mini model at 720p and re-render only the keeper on the top model.
Where Cast fits
- You want to stop doing token arithmetic: Cast shows each video's price in credits on the button, and a 5-second Standard video costs 52 credits.
- You want the draft-cheap pattern built in: Standard runs on Seedance 2.0 Fast at 720p, and Pro (Seedance 2.5, 1080p) is there for keepers on Premium.
Where it does not:
- The lowest per-second price at volume: an API at list price will always beat an app with a margin.
- Long or 4K renders: Cast stops at 15 seconds and 1080p.
People also ask
- Why is Seedance 2.5 so much more expensive than 2.0 Fast?
- It is a bigger model and it goes to 1080p. At 720p BytePlus lists 2.5 at 10.70 USD per million tokens and 2.0 Fast at 5.6[1]; at 1080p the token count rises another 2.25 times. Compare the two in Seedance 2.0 Fast vs 2.5.
- What costs the most in an AI video?
- Resolution and length, then the model tier. Video inputs and generated audio add on some platforms, for example Replicate's Seedance 2.5 costs about four times more per second with a video input[5].
- Why do AI images cost so little by comparison?
- An image is one frame. A 5-second clip at 24 frames a second is 120 frames that must also stay consistent with each other, so it needs far more compute.
Sources
Claims about other companies come from their own pages, read on the date shown. Products and prices change; check the link for the current version.
- BytePlus ModelArk: model pricing (read )
- Gemini Apps Help: usage limits (read )
- Google: Gemini API pricing (read )
- fal.ai: Seedance 2.5 text-to-video (read )
- Replicate: bytedance/seedance-2.5 (read )
- Google Cloud: generative AI pricing (read )
- Runway: pricing (read )
- Wan2.2 on GitHub (read )
Related answers
- How much does AI video cost per second in 2026?
- What is the cheapest AI video generator?
- How do AI video credits work?
- Seedance 2.0 Fast vs Seedance 2.5: which should I use?
- How much does Seedance cost per video?
Part of the Cast answers library (Prices and costs).
Updated