How to make AI videos longer: two ways past Grok’s 15-second and Kling’s 30-second walls
There are two ways to make an AI video longer: an extension feature, or starting the next clip from the last frame. In one run, Kling 4.0, Seedance 2.5 and Wan 3.0 reach 30 seconds. Sora 2 stops at 20, Grok at 15 and Veo 3.1 at 8. The Grok app extends videos to 30 seconds in total. When we asked the three 30-second models for 31 seconds in free API estimates, all three refused (checked Oct 26, 2026).
As of Oct 26, 2026, the longest an AI video model will go in one generation is 30 seconds. Kling 4.0, Seedance 2.5 and Wan 3.0 all reach it. Sora 2 stops at 20 seconds, Grok Imagine 1.5 at 15, Gemini Omni Flash at 10 and Veo 3.1 at 8. Anything longer means extending a clip or starting the next one from the last frame of the one before. The Grok app extends videos to 30 seconds in total.
The lengths and extension rules below come from each company’s API docs and the Grok app’s official pages. Sera Studio added what the APIs actually accepted, checked with free price estimates and public schemas.
How long one AI video run can be
Kling/Seedance/Wan
- Longest single run
- 30 sKling 4.0, Seedance 2.5, Wan 3.0
- How to extend
- Wan extends both ways. Kling plans 2-minute extensions
- Docs
- Kling, BytePlus, Alibaba
Sora 2
- Longest single run
- 20 sPick 4, 8, 12, 16 or 20 s
- How to extend
- Up to 20 s, 6 times, 120 s total
- Docs
- OpenAI
Grok
- Longest single run
- 15 sImagine 1.5, app at 720p
- How to extend
- App: 30 s total. API: 2–10 s at a time
- Docs
- xAI, Grok app
Omni, Veo 3.1
- Longest single run
- 10 · 8 sOmni 3–10 s, Veo 4, 6 or 8 s
- How to extend
- Omni to 40 s. Veo 7 s at a time to 37 s
- Docs

Three models make 30 seconds in one run
- Kling 4.0 runs 3 to 30 seconds. Kling’s release notes say it generates up to 30 seconds at once and takes up to 10 keyframes (Meet All-New Kling 4.0). The faster 4.0 Flash runs 3 to 20 seconds.
- Seedance 2.5 runs 4 to 30 seconds. BytePlus gives the same range in its video generation tutorial and on the Seedance 2.5 page, at 480p, 720p or 1080p and 24 fps.
- Wan 3.0 runs 2 to 30 seconds. Alibaba Cloud’s guide says that setting the duration to -1 lets the model pick a length (Wan3.0 Video Generation).
Sora 2 tops out at 20 seconds, Grok at 15, Omni at 10 and Veo at 8
- Sora 2 takes 4, 8, 12, 16 or 20 seconds, with 4 as the default. OpenAI raised the maximum from 12 to 20 seconds in March 2026 (Sora 2 Prompting Guide).
- Grok Imagine Video 1.5 (grok-imagine-video-1.5) takes 1 to 15 seconds, 8 by default. Reference-to-video on grok-imagine-video, the model without a version number, is capped at 10 seconds (Video Generation).
- Gemini Omni Flash makes 3 to 10 seconds (Gemini Omni Flash).
- Veo 3.1 makes 4, 6 or 8 seconds, and only 8 when it starts from reference images (Veo 3.1).
The Grok app extends videos to 30 seconds in total
- xAI’s Grok page puts text-to-video in the Grok app at up to 15 seconds in 720p (Grok).
- Grok added video extensions on Feb 17, 2026, according to its release notes. You set the added length from a banner and the toolbar, and the whole video caps at 30 seconds (Grok Release Notes, Feb 17, 2026).
- According to the Oct 3 release notes, you can choose whether an attached image becomes the first frame, last frame, a loop or a reference. So the app can also start a new clip from the last frame of the previous one.
- The API’s extension rules differ from the app’s. They’re in the next section.
Extensions feed a finished video back in and continue it
An extension feeds a finished video back to the model and generates what comes next. Some models can also add footage before it.
- Sora 2 extensions use the whole source clip as context and return a new stitched video. Each one adds up to 20 seconds, a video can be extended six times and the whole thing caps at 120 seconds. Extensions take only the source video and a prompt, with no characters or image references (Video generation with Sora).
- Omni Flash adds 10 seconds at a time, up to a total length of 40 seconds. It uses the last 10 seconds of the original as context and edits some of its final frames to hide the seam (Generate and edit videos with Gemini Omni Flash).
- On Google Cloud (Gemini Enterprise Agent Platform), Veo 3.1 extends a video by 7 seconds at a time. The input has to be 1 to 30 seconds long at 24 fps, in 720p, 1080p or 4K and in 9:16 or 16:9. Extended Veo videos run up to 37 seconds (Extend videos).
- Grok extensions run on the grok-imagine-video model. xAI’s docs say it continues from the source’s last frame and don’t say whether it reads the whole clip. The source must be an MP4 of 2 to 15 seconds, and the extension is 2 to 10 seconds, 6 by default. The output keeps the source’s aspect ratio and resolution, capped at 720p. The duration setting covers only the new part (Video Extension).
- Wan 3.0 extends forward, backward or both when you pass the source video with a prompt that asks to extend or continue it. Setting the ratio to adaptive keeps the source’s aspect ratio.
- Kling’s 4.0 release notes say that uploading a video and extending it several times, up to 2 minutes, is coming soon.
| Model | Per run | How to extend (docs) |
|---|---|---|
| Kling 4.0 | 3–30 s | Several times up to 2 min (coming soon) |
| Seedance 2.5 | 4–30 s | Not found in the docs |
| Wan 3.0 | 2–30 s | Forward, backward or both |
| Sora 2 | 4, 8, 12, 16, 20 s | Up to 20 s, 6 times, 120 s total |
| Grok Imagine 1.5 | 1–15 s | grok-imagine-video, 2–10 s on a 2–15 s source |
| Grok app | Up to 15 s, 720p | Up to 30 s total |
| Omni Flash | 3–10 s | 10 s at a time, 40 s total |
| Veo 3.1 | 4, 6, 8 s | On Google Cloud, 7 s at a time to 37 s |
Without an extension feature, start again from the last frame
A model that can turn an image into video can still build a longer piece without an extension feature. You use the last frame of one clip as the start image of the next.
OpenAI’s docs say when to use each. Use an extension to preserve motion, camera direction and scene continuity. Use input_reference if you only need to set the opening frame.
You can also make several short clips and join them in an editor. OpenAI’s prompting guide says the model follows instructions more reliably in shorter clips, and that two 4-second clips stitched together may work better than one 8-second clip.
In price estimates, five models accepted only their published lengths
On Oct 26, Sera Studio checked length limits with free Higgsfield API estimates and fal’s public schemas. An estimate asks for a price and generates nothing. We didn’t run Sora 2, Omni or Veo through estimates.
- Kling 4.0 at 31 seconds was rejected as over the 30-second maximum. 30 seconds passed, and 2 seconds was rejected as under the 3-second minimum.
- Seedance 2.5 and Wan 3.0 were also rejected at 31 seconds, and below their minimums at 3 and 1 seconds.
- Kling 3.0 and Grok Imagine 1.5, both 15-second models, were rejected at 16.
- Higgsfield’s API has a separate Seedance 2.5 extension route that adds 4 to 30 seconds. Its estimate text says both the input and the generated duration are billed.
- fal’s MiniMax H3 Max extension adds 0.71 to 15 seconds to a source of 1.625 to 60 seconds. The new clip also includes the source’s last 1.625 seconds.
| Model, where checked | Accepted | Extension |
|---|---|---|
| Kling 4.0, Higgsfield estimate | 3–30 s. 2 s and 31 s rejected | None, only a video edit route |
| Kling 3.0, Higgsfield estimate | 3–15 s. 16 s rejected | None |
| Wan 3.0, Higgsfield estimate | 2–30 s. 1 s and 31 s rejected | None |
| Seedance 2.5, Higgsfield estimate | 4–30 s. 3 s and 31 s rejected | 4–30 s. The source length is billed too |
| Grok Imagine 1.5, Higgsfield estimate | 1–15 s. 16 s rejected | None |
| Grok Imagine, fal schema | - | 2–10 s on a 2–15 s source |
| Veo 3.1, fal schema | 4s, 6s, 8s | Default 7s, 720p or 1080p source |
| MiniMax H3 Max, fal schema | 0.92–15 s | 0.71–15 s on a 1.625–60 s source, overlapping its last 1.625 s |
Which way to go
The rest of this section is our analysis, not the companies’ documentation.
Two 30-second generations joined together give you a one-minute Short or Reel. The longer each run, the fewer seams you have to hide.
Extensions keep the action continuous, but they can cost more. BytePlus’s price list counts the length of any input video in Seedance 2.5 tokens. By that rule, extending a 20-second 720p clip by 10 seconds bills 30 seconds of tokens, the 20 you put in plus the 10 new ones, at $6.40 per million. A fresh 10-second clip with no video input bills 10 seconds at $10.70 per million. Assuming the same tokens per second, one extension costs about 1.8 times as much as a new 10-second clip.
Whether you get back the whole video or only a new piece also varies. Sora 2 returns the stitched result. MiniMax H3 Max returns a new clip that overlaps the last 1.625 seconds of the source. Omni edits the source’s final frames, so the end of your original file may not match the result. Check for overlap before you join clips in an editor.
Starting from the last frame works with any model that turns images into video, but the next clip sees only that one frame. Motion direction and sound can break at the join, which is easier to hide at a cut between scenes.
What we couldn’t confirm
- Kling hasn’t given a release date for 2-minute extensions.
- We didn’t find Seedance 2.5 extension length rules in BytePlus’s docs. We confirmed the extension route on Higgsfield’s API.
- Alibaba’s guide caps reference videos at 15 seconds in total, while Higgsfield’s Wan 3.0 reference field allows 30. Neither states a maximum source length for extensions.
- We didn’t check whether Sora 2, Omni, Veo or Grok extensions bill the input video’s length.
- The official pages we read don’t say whether Grok app lengths differ by plan.
- The Sora and Kling apps may offer different options from their APIs. This piece follows the API docs, the Grok app’s official pages and our own API checks.
- Sera Studio’s Motion Studio uses most of these models. Its Continue option passes the source video to Seedance models and gives the rest the last frame as their start image. Joining scenes into one export costs no credits. Sera Studio wrote this piece.


