Why are AI videos only 8 to 15 seconds, and how do I make longer ones?
Clips are short because every extra second costs more compute and gives the model more time to drift off the character, so most models cap one generation at 8 to 15 seconds and the newest at 30. To go longer, chain clips: render the first part, take its last frame, and start the next part from that frame, or use a model's built-in extend feature.
How long each model goes in one generation (as of 2026-10-18)
| Model | Longest clip | Longer with extensions? |
|---|---|---|
| Google Veo 3.1, Fast and Lite[1] | 8 s (4, 6 or 8) | Veo 3.1 extends its own clips by 7 s, up to 20 times[1] |
| Seedance 2.0, 2.0 Fast, 2.0 Mini[2] | 15 s | Not stated in BytePlus's docs; we chain clips |
| Seedance 2.5[3] | 30 s | Multi-round extensions, per ByteDance[3] |
| Kling 3.0 / Kling 4.0[4] | 15 s / 30 s | Not stated on the page we read |
| Midjourney video[5] | 5 s | 4 s at a time, up to 4 times: 21 s[5] |
| MiniMax Hailuo 2.3 (API)[6] | Priced per 6 s video | Not stated |
Two reasons keep the numbers small. Cost first: providers bill video by output size and seconds, and BytePlus prices Seedance in tokens of width × height × frame rate × seconds ÷ 1024[7], so a 30-second clip costs about six times a 5-second one. Then drift: the longer a single generation runs, the more a face, an outfit or a room has time to change. Short clips that start from a fixed picture keep a character on-model far better, which is why our character consistency guide builds long pieces from many short clips.
OpenAI's Sora 2 API once allowed generations to be extended to 120 seconds in total[9], but OpenAI shut that API down on 2026-09-24[8].
How to chain clips without a visible jump
This is what worked for us on Seedance, making 30-second videos out of 15-second clips:
- Write the whole video as one timeline first (0-4 s, 4-9 s and so on), then split it at a calm moment, not mid-gesture.
- Render part A. Export its very last frame as an image.
- Render part B as image-to-video from that frame, with the second half of the timeline as the prompt. Repeat the character description word for word.
- Do not feed part A back in as a video reference for part B. In our test the model restarted the scene and there was a visible jump at the 15-second mark.
- Join the parts with a frame-accurate cut and a short audio crossfade (we used 80 ms). If one ending is bad, cut at its last clean frame and render a short clip from exactly that frame.
Our 30-second dance Own It was made exactly this way, as two 15-second Seedance 2.0 Fast halves, at about $2.50 per take.
Does a longer clip cost more?
Yes, almost linearly. Atlas Cloud's price endpoint quoted our account $0.29 for a 5-second Seedance 2.0 Fast clip, $0.58 for 10 seconds and $0.87 for 15. A single 30-second Seedance 2.5 clip quoted $9.03 at 720p and $17.77 at 1080p. Two chained 15-second Fast halves cost far less than one 30-second 2.5 take, at the price of a seam you have to hide. Use one long take only when continuity matters more than money, for example a single uninterrupted dialogue shot.
Talking-avatar tools are the exception to short clips because they animate a face over an audio track rather than inventing a whole scene: Hedra says its Character 3 model makes talking and singing video up to 10 minutes long[10]. That only helps if your video is a person talking to camera.
Where Cast fits
- Short scenes of a recurring character: Cast makes 5, 10 and 15-second vertical clips and keeps the same character reference in every one, so several clips cut together still look like one person.
- Testing a beat before you pay for a long render: a 5-second Standard clip costs 52 credits and your first one is free.
Where it does not:
- Anything longer than 15 seconds in one piece: Cast has no extend button or timeline editor to join clips.
- Multi-episode films or series: Katechon Arc (https://arc.katechon.technology/) is built for that.
- Long talking-head videos read from your own audio.
People also ask
- Why does Veo stop at eight seconds?
- Google's Veo 3.1 models generate 4, 6 or 8 second clips, and 1080p, 4K, reference images and extension all require the 8-second length[1]. Veo 3.1 can then extend a clip it made by 7 seconds at a time, up to 20 times[1].
- Which AI video models make 30 seconds in one go?
- Seedance 2.5 goes up to 30 seconds per generation[3], and Kling says Kling 4.0 supports clips from 3 to 30 seconds[4]. Seedance 2.0 and Kling 3.0 stop at 15 seconds[2][4].
- How do I stop the character changing between clips?
- Start every clip from a still frame, either the last frame of the previous clip or a frame painted from the same master picture, and repeat the same written description each time. Our consistency guide covers the details.
- Can I just loop a short clip?
- You can, but platforms may not reward it: TikTok says content made only of looping videos does not count as original for its Creator Rewards Program[11].
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.
- Google: Veo on the Gemini API (read )
- BytePlus ModelArk: Seedance 2.0 (read )
- ByteDance Seed: introducing Seedance 2.5 (read )
- Kling AI: AI video generator (read )
- Midjourney docs: Video (read )
- MiniMax: API pricing overview (read )
- BytePlus ModelArk: model pricing (read )
- OpenAI: Sora 2 model page (read )
- OpenAI: video generation guide (Sora API) (read )
- Hedra: models (read )
- TikTok: Creator Rewards Program (read )
Related answers
- How do I keep an AI character consistent across videos?
- How much does AI video cost per second in 2026?
- Seedance 2.0 Fast vs Seedance 2.5: which should I use?
- How do I write a good Seedance prompt?
- What is Kling 3.0?
Part of the Cast answers library (Other tools and models).
Updated