AI Video Generator Long Videos: Chaining Past the Cap

Every AI video model caps a single generation well under a minute. Seedance 2.5 stops at 30 seconds, Seedance 2.0 and MiniMax H3 at 15, and Veo 3.1 is fixed at 8.^[1]^[2]^[3]^[4] There is no setting that lifts those ceilings, and no AI video generator long videos feature hiding behind a paid tier. The limit is architectural.

So every multi-minute AI video you have seen was assembled from segments. The question worth answering is not which model makes long videos, it is how to join segments without the join being visible, what that costs when you count the segments you throw away, and which parameter combinations quietly become illegal the moment you start chaining. This covers all three, using the API rather than a timeline editor.

TL;DR

30 seconds is the current single-call ceiling, on Seedance 2.5.^[1] Every AI video generator long videos workflow past that point is chained segments.
return_last_frame is the joint. Ask for it, then feed output.last_frame_url into the next call as the first frame.^[1]
Chaining forces size: adaptive. A first-frame job rejects any other aspect ratio at submit, so pick your framing in segment one.^[1]
A chained segment cannot also carry a reference video. image_with_roles is mutually exclusive with video_urls and audio_urls.^[1]
A three-minute 720p piece is about $48 of generation if nothing is rejected, and closer to $144 at three attempts per segment.^[5]

Why an AI video generator long videos mode does not exist

Published limits, from each model's parameter reference:

Model	Single-call duration
Seedance 2.5	4–30 s, or `-1` to let the model choose^[1]
Seedance 2.0	4–15 s^[2]
MiniMax H3	4–15 s, default 6^[3]
Grok Imagine 1.5	1–15 s^[6]
Veo 3.1	fixed at 8 s upstream^[4]

Thirty seconds is the best available, and it is recent. Treat any AI video generator long videos claim as either chaining on your behalf or describing something other than one continuous generation.

That distinction matters for cost. A tool that chains for you still pays for every segment, and so do you.

The chain: last frame out, first frame in

The mechanism is two fields. Set return_last_frame: true on a generation and the completed task carries output.last_frame_url alongside the video.^[1] Hand that URL to the next call as a first frame and the second clip starts on the exact pixel the first one ended on.

Segment one, establishing the framing for the whole piece:

curl https://reapi.ai/api/v1/videos/generations \
  -H "Authorization: Bearer rk_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-2.5-face",
    "prompt": "Wide shot, harbour at dawn, slow push in",
    "resolution": "720p",
    "size": "16:9",
    "duration": 30,
    "return_last_frame": true
  }'

Segment two onward, starting from the previous frame:

curl https://reapi.ai/api/v1/videos/generations \
  -H "Authorization: Bearer rk_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-2.5-face",
    "prompt": "The camera continues past the moored boats toward the market",
    "resolution": "720p",
    "duration": 30,
    "return_last_frame": true,
    "image_with_roles": [
      { "url": "https://cdn.reapi.ai/media/tasks/.../0.png", "role": "first_frame" }
    ]
  }'

Note what disappeared from the second request. There is no size.

Three rules that only bite once you chain

Chained segments must run adaptive. A request carrying image_with_roles accepts only size: "adaptive" or no size at all; anything else is rejected at submit rather than minutes later.^[1] This is not arbitrary, the output follows the first frame's geometry, which is exactly the behaviour continuity needs. The practical consequence is that segment one decides the aspect ratio for the entire piece, so get it right before you generate 6 more.

A chained segment cannot carry reference video or audio. image_with_roles is mutually exclusive with both video_urls and audio_urls.^[1] If a mid-sequence shot needs a motion reference, that shot cannot also be frame-chained. Pick one per segment.

image_urls and image_with_roles cannot be combined. Character reference images and frame chaining are separate modes.^[1] For a chained sequence with a recurring character, carry the character through the frame itself rather than through reference images.

Those three constraints are enforced before submit, which is the good case. You get a 400 in a second rather than a failed task in four minutes.

Where chains actually break

Frame continuity is not scene continuity. The joint is pixel-perfect and the drift shows up somewhere else.

Colour and exposure walk. Each segment re-derives its own grade from one frame. Over six segments the sequence can wander perceptibly even though every individual join is seamless. Grade the assembled cut, not the segments.

Motion discontinuity at the seam. A still frame carries no velocity. If segment one ends mid-pan, segment two starts from a static frame and has to re-establish the movement, which reads as a stutter. Write segment endings that come to rest, or accept a cut on the joint and treat it as a shot change rather than a continuation.

Audio does not chain. Each generation synthesises its own track. Six segments give you six independent audio beds, and the seams are more audible than the visual ones. For anything longer than about a minute, generate without audio and lay a single track over the assembled cut.

Identity drift on people. A face reconstructed from one frame each time accumulates error. This is where a chained sequence usually gives itself away first.

None of these are fixable by prompting harder. They are properties of restarting a generation from a still image, and the practical answer is to design the piece so its cuts land where the seams are.

What three minutes actually costs

180 seconds at 720p on Seedance 2.5 is six 30-second segments. At the published rate of $0.266824 per second:^[5]

Attempts per segment	Total generated seconds	Cost
1 (nothing rejected)	180	$48.03
2	360	$96.06
3	540	$144.09

At 480p the same three tiers are $21.35, $42.70 and $64.04.^[5]

The realistic number is the second or third row. Chained work has a failure mode single clips do not: a segment can be individually good and still be wrong, because it drifted from the one before it, and you only find out after generating it. Budget the retries or you will be surprised twice.

An obvious saving follows. Draft the whole sequence at 480p, confirm the chain holds and the cuts land, then regenerate only the segments that survive review at 720p. That is $21 of drafting instead of $48, and it is the single biggest lever on a long piece.

When to let the model pick the length

Seedance 2.5 accepts duration: -1, which hands the choice to the model, and it is required rather than optional for video-edit prompts.^[1] Billing reserves at the 30-second cap and settles to the length actually produced, refunding the difference.^[5]

For chaining, explicit durations are usually better. A sequence built from known-length segments is one you can cut to music or to a script; a sequence of model-chosen lengths is one you have to conform afterwards. Use -1 for edits, where the vendor requires it, and for single clips where the natural length is genuinely unknown.

FAQ

What is the longest single AI video generation available?

30 seconds, on Seedance 2.5.^[1] Most other models cap at 15 and Veo 3.1 is fixed at 8.^[2]^[3]^[4]

How do I make an AI video generator produce long videos?

Chain generations. Set return_last_frame: true, then pass the returned last_frame_url into the next call as a first_frame role.^[1] There is no single call that produces minutes of output.

Why does my chained request get rejected for aspect ratio?

Because a first-frame job only accepts size: "adaptive", or no size field at all.^[1] Remove size from every segment after the first.

Can I use reference images and frame chaining together?

No. image_urls and image_with_roles are mutually exclusive, as are image_with_roles and video_urls.^[1]

What does a three-minute AI video cost?

About $48 at 720p if every segment lands on the first attempt, and roughly $144 at three attempts each.^[5] Drafting at 480p first cuts the exploration phase to about $21.

Why does the audio jump between segments?

Each generation makes its own audio. Turn generate_audio off for chained work and lay one track over the finished cut.

Do the joins look seamless?

The joins do. The drift between segments is what gives a chain away: colour walk, restarted motion, and identity error on faces.

Building length out of segments

There is no AI video generator long videos setting, and treating its absence as a limitation of the tool rather than of the technology leads people to buy plans looking for it. What exists is a 30-second ceiling, a frame-chaining mechanism, three parameter rules that are enforced at submit, and four kinds of drift that are not.

Plan around them in that order. Set the aspect ratio in segment one because you cannot change it later. Write segment endings that come to rest. Generate silent and score the assembly. Draft at 480p and finish at 720p. Then price the whole thing at three attempts per segment, because that is what it will actually take.

Do that and an AI video generator long videos workflow stops being a feature you are shopping for and becomes a build you can budget. Parameter reference and error codes are at reapi.ai/docs/seedance-2-5, and current per-second rates are on the Seedance 2.5 model page.

References

reAPI. doubao-seedance-2.5-face — duration, return_last_frame, image_with_roles and cross-field constraints. Retrieved August 2026 from reapi.ai/docs/seedance-2-5
reAPI. Seedance 2.0 — parameter reference and duration limits. Retrieved August 2026 from reapi.ai/docs/seedance-2-0
reAPI. MiniMax H3 — parameter reference and duration limits. Retrieved August 2026 from reapi.ai/docs/minimax-h3
reAPI. Veo 3.1 — parameter reference, fixed 8-second duration. Retrieved August 2026 from reapi.ai/docs/veo3-1
reAPI. Seedance 2.5 — published per-second rate band and auto-duration settlement. Retrieved 9 August 2026 from reapi.ai/models/seedance-2-5
reAPI. Grok Imagine 1.5 — parameter reference and duration limits. Retrieved August 2026 from reapi.ai/docs/grok-imagine-video-1-5