Best AI for video editing: does the most expensive model win?

The most expensive AI model is not the best at editing a video. We edited the same clips with the four Endless models that can swap someone inside an existing video. Seedance 2.5, the most expensive per second, keeps the shot and the sound very well, but with people it never fully replaces the face. Wan 3.0 Prime was the best on identity and on fidelity to the clip, and Kling 4 came very close for fewer credits. The cheapest, MiniMax H3, remakes the clip instead of editing it.

It all started with Francisco, our CEO, who took the celebration after Enzo Fernández's stunning goal against England, uploaded photos of Ponyo the cat and asked Seedance 2.5 to replace number 24 with Ponyo. What came out was seven seconds of a calico cat in the number 24 jersey, paws behind its ears, hugged by its teammates and saluting at the end.

By noon, everyone in the office with a cat was trying the same thing. And the usual questions started: why is my cat missing the jersey? Why does mine have human arms? Why did the commentary change?

Seedance 2.5 had been our default for editing video, and Ponyo backed that up. It is also the most expensive of the four models, so the wave of cats gave us an excuse to ask something broader: does paying more for a model get you a better edit? We set up a test bench with the four Endless models that edit an existing video: Seedance 2.5, Kling 4, Wan 3.0 Prime and MiniMax H3.

How we tested

We ran three video-to-video editing tests, all at 720p and all in the same Endless workspace:

The process was the same for all of them.

1. Pick the clip

Clips between 7 and 10 seconds, with clear actions and, in the celebration, a cut between shots. The same file went to all four models.

2. Build the references

Each test has its own references, and we show them in each section:

• Dancers: a character sheet of a made-up character, generated in Endless with Nano Banana 2.1 for 13 credits.
• Cats: two photos of the cat, one of the face and one of the body.
• Booth: a front-facing photo of me. From that photo we built a character sheet in Nano Banana Pro for 25 credits, and then we tried a 2-second selfie video.

3. One prompt per model

The same request is not written the same way for all four. Each model understands a different structure best, marks timing in its own way and names references in its own style. To find the best prompt for each one we used their prompting guides, the same ones Enhance prompt uses in Endless: you write the request once, pick the model and Enhance prompt rewrites it in the structure that model understands best.

Here, for example, is the last shot of the celebration in each one:

Kling 4        6.6-6.9s: hard cut to a side view on the pitch. The same
               cat, still wearing the navy number 24 jersey, raises one
               paw to its ear in a salute.

Seedance 2.5   TIMELINE: … 6.6-6.9s hard cut to a side view on the pitch:
               the SAME cat, STILL wearing the number 24 jersey, raises
               one paw to its ear in a salute.

MiniMax H3     [6.6s] Hard cut to a side view on the pitch: the cat, in
               the number 24 jersey, raises one paw to its ear in a salute.

Wan 3.0 Prime  …in the final side-view shot the cat itself, wearing the
               number 24 jersey, raises one paw to its ear in a salute.
               Negative: human face, human arms, tattoos, cat without
               jersey…

4. Generate

We ran the four models on the same clip, with the same references and at 720p, and put the results side by side on the Endless canvas.

5. Compare

The canvas is the easiest way to compare: you play the four videos side by side and the differences jump out right away. For this post we also made one video per test with the original, the reference and the four results in sync, and pulled frames from every video at the same seconds, one row per model. In the tests with people we also cropped the face at the same moment. For each result we looked at four things:

• Identity: whether it is really that cat or that person.
• Fidelity: whether the clip is still the same, with the same shots, gestures and timing.
• Audio: whether it keeps the original sound.
• Credits: what each attempt costs.

How much does it cost to edit a video with each model?

Each model has its own price in credits. For a 10 s clip at 720p:

These are Endless prices as of October 28, 2026.

Seedance costs the most per second: more than double H3 and almost 50% more than Kling. If price measured quality, Seedance should win every test.

Dancers: swapping one person without touching the other

The four models, in sync with the original and the reference.

Kling and Wan come out on top. They swap in the character, follow the choreography and leave the other dancer and the music untouched. H3 looks the most like the character, but it remakes the clip: it changes how both dancers move and generates new music. It behaves more like "take inspiration from this video" than "edit this video".

Seedance respects the clip and the music, but the character never quite becomes the one on the sheet: it drops the glasses and slims the face.

Cat in the celebration: short shots and one last cut

This clip has a catch: it ends with a 0.3 s side shot where number 24 salutes. That is why each model's prompt describes both shots with their timing, names that last shot, asks for cat paws in place of the player's arms and puts the audio on a separate line.

The four models, in sync with the original and the cat photos.

Seedance and Wan get everything right: the cat is in both shots, wearing the jersey, it does the final salute, and both keep the timing and the scoreboard. Seedance's cat is almost identical to the one in Francisco's video. The difference is the price: 298 credits against Wan's 244.

Kling treats the short shot as a new take and drops the cat's jersey there. H3 moves the hug earlier and skips the last shot.

This applies to any model: if a shot lasts less than a second, name it in the prompt with its timing.

Replace a person in a video with AI: me at a booth

To finish we went back to a person. The clip is one continuous 10 s shot: someone at a booth turns around, talks to the camera and holds up a laptop.

First we built a character sheet with Nano Banana Pro from a single photo: two full-body views and three portraits. It cost 25 credits. Then we used it as the reference in all four models.

With the character sheet: the four models, in sync with the original.

Then we tried a video of me as the reference instead of the character sheet: a 2-second selfie, together with the same front-facing photo.

The 2-second reference selfie.

With the reference video: the four models, in sync with the original.

The reference video gives Wan details the character sheet did not have, like the real eye color and the volume of the hair. Kling comes very close and keeps all of the audio. Seedance still does not fully replace the face, with the sheet or with the video, though it keeps the shot and the sound very well.

Character sheet or reference video?

For locking in a character, the character sheet had been our go-to tool: a single image with the character from the front, the side and the back, which the model uses as a reference in every shot. It works, and in these tests we used it for the dancers' character.

The booth test gave us three reasons to prefer a video when the person is real. A character sheet is an image generated from a photo, so it makes up whatever the photo does not show; with a 2-second selfie, Wan picked up the real eye color and the volume of the hair, which the sheet did not have. Also, in a video the face turns, talks and changes expression while staying the same face, and that shows in a clip where the presenter talks, turns around and looks back at the camera. And you save a step: there is no sheet to generate and check before editing. Recording a few seconds on your phone and uploading them to Endless is enough.

The character sheet still has its place: it is how you give a body to a character who does not exist, like the one in the dancers test. When the character is real, though, we go straight to video more and more often.

Seedance vs Kling vs Wan vs MiniMax: what is the best AI for video editing?

To answer the opening question we weighted the four criteria by what matters most to us when we edit: that it is that person or that cat, that the clip stays the same, that the audio holds, and what it costs to get there. Scores go from 1 to 5 and sum up all the tests.

Wan 3.0 Prime wins clearly. It is the most faithful on identity and on the clip, and it almost always keeps the audio. It is not cheap, but each attempt is usually usable.

Kling 4 is the most efficient alternative: 17% cheaper than Wan and very close in quality. Its weak spot is very short shots, where it can lose details like the jersey.

Seedance 2.5 keeps the shot and the sound very well, and it does an excellent job with the cat in the celebration. With people, though, it never fully swaps the face. And it is the most expensive.

MiniMax H3 is by far the cheapest, but it remakes the clip instead of editing it. It works if you want a new version inspired by the original. If you want the same clip with a different lead, it is the wrong pick.

What we learned running them in Endless

• One prompt per model, with Enhance prompt. The same request written the same way for all four does worse. Enhance prompt rewrites it in each model's structure: Edit @Video1. and timed segments for Kling, blocks in caps for Seedance, [6.6s] for H3 and a Negative: line for Wan.
• Name short shots with their timing. A shot under a second that is not in the prompt is the first thing to go.
• For a real person, use a video. A few seconds recorded on your phone give the model detail and consistency that a character sheet lacks. Save the sheet for characters who do not exist.
• Compare on the canvas. With the four results side by side, differences in identity, timing and audio show up in seconds.

So, is the most expensive model the best for editing video?

No. Seedance 2.5, the most expensive, came third, and the cheapest does not work either: MiniMax H3 costs less than half of Wan, but it makes a different video. For editing, the best option was Wan 3.0 Prime, the second most expensive, and Kling 4 comes very close if you want to spend less.

We stopped using Seedance 2.5 by default, even though Ponyo, the cat that started it all, is among the best results of all the tests. With people it does not swap the face, and each attempt costs between 20% and 50% more than on Wan or Kling. For fewer credits, Wan makes an equally good cat and leaves you room to iterate.

You can run the same comparison with your own clips: upload the video and the references to Endless, write the request once, run it through Enhance prompt for each model and watch the results side by side on the canvas. When the next model comes out, run the same inputs again and compare.