We ran the same 6 prompts 3 times on 8 video models. Counting only the clips worth keeping, a 5-second clip cost $0.36 on MiniMax H3 and $2.30 on Seedance 2.5, about 6× apart. Usable rates were close (81%–100%), so price decided most of it.
Every one of the 144 clips is below with its score. The scores come from an AI model, not people, so watch the clips and judge for yourself. VisionMorphic sells all eight models; the prices are ours.
Cost per usable clip
Text to video, 5 seconds, 720p (MiniMax H3 starts at 768p), 16:9, with sound, 18 clips per model. A usable clip scores 1 and a borderline one 0.5. Credits per usable clip = credits per clip ÷ usable rate, because you pay for the clips you throw away too. Dollars are at the Basic plan rate ($9.90 for 600 credits); credit packs cost more per credit and larger plans less. Median wait is how long a clip took to finish in our runs on November 2, 2026 and November 6.
Price per clip, not failure rate, set the order. MiniMax H3 is the cheapest per clip and stays the cheapest per usable clip even at 92% usable; Seedance 2.5 costs 6× as much per clip and stays the most expensive.
With 18 clips per model, one clip moves the usable rate by about 6 points. Wan 3.0 scored best at 100%, but most of the other gaps are one or two clips.
Hands were the hardest shot: 79% usable across all models, against 92%–98% for the other five. The usual problem was splitting the orange with much of the peel still on, often into halves that look cut with a knife.
Seedance 2.0 Mini, MiniMax H3 and Wan 3.0 peeled and split the orange cleanly in all 3 runs; Kling 3.0 Omni did not manage it once.
Kling 3.0 Omni misspelled the sign in 2 of 3 runs ("OPEN 4 HOURS", "OPEN 24 HURS"). Every other model spelled it right every time.
Seedance 2.0 put a drone in the frame in all 3 fast camera runs. The prompt says "FPV drone shot", meaning the camera is the drone, and no other model read it the other way. With Seedance 2.0, leave the word "drone" out.
All 24 dialogue clips said the line; the transcripts matched "I already ordered for you".
Usable clips by shot
Out of 3 runs per shot. Bold marks a shot where a model lost points.
Shown as generated, with no editing. Click a clip to play it with sound. Every clip that lost points says why.
Full-body walk
Tests legs and gait over several steps. 96% usable across all models.
Prompt used for every model
A woman in a long red coat walks toward the camera along a busy city sidewalk on a sunny afternoon. Full body in frame, natural relaxed gait, arms swinging. Steady tracking shot moving backward with her. Realistic, natural light. Sound: footsteps and street ambience.
Tests fingers, object handling. 79% usable across all models.
Prompt used for every model
Close-up of a person's hands peeling a mandarin orange over a wooden table, pulling the peel away in a few pieces, then splitting the fruit into two halves. Fingers clearly visible, natural movement, five fingers on each hand. Realistic, soft window light. Sound: peel tearing, quiet room.
Run 1 · BorderlinePeels only part of it, then splits the orange with most of the peel still on.Run 2 · UsableRun 3 · Usable
Kling 3.0 Omni
1.5 / 3 usable
Run 1 · BorderlinePeels only part of it, then splits the orange with most of the peel still on.Run 2 · BorderlineTakes off a piece or two of peel, then splits the orange with the rest still on.Run 3 · BorderlineTakes off a piece or two of peel, then splits the orange with the rest still on.
Run 1 · UsableRun 2 · BorderlinePeels only part of it, then splits the orange with most of the peel still on.Run 3 · Usable
Seedance 2.0
2 / 3 usable
Run 1 · BorderlinePeels only part of it, then splits the orange with most of the peel still on.Run 2 · UsableRun 3 · BorderlinePeels only part of it, then splits the orange with most of the peel still on.
Run 1 · BorderlinePeels only part of it, then splits the orange with most of the peel still on.Run 2 · Not usableStill peeling when the clip ends; the orange is never split.Run 3 · Usable
Face turning
Tests identity stays the same through a profile-to-front turn. 92% usable across all models.
Prompt used for every model
Medium close-up of a middle-aged man with a short grey beard sitting by a window, seen in profile. He slowly turns his head to face the camera and gives a small smile. His face stays the same person throughout. Realistic, warm window light, shallow depth of field. Sound: quiet room tone.
Run 1 · UsableRun 2 · UsableRun 3 · BorderlineStarts at a three-quarter view, so the turn is small.
Readable text
Tests spelling and stability of on-screen text. 92% usable across all models.
Prompt used for every model
A neon sign on a dark brick wall flickers on and reads exactly "OPEN 24 HOURS" in bright pink letters. Light rain falls in front of it and the wet wall reflects the glow. Slow push-in. Sound: neon buzz and rain.
Tests coherence under fast motion. 94% usable across all models.
Prompt used for every model
FPV drone shot racing low and fast through a pine forest, weaving between tree trunks, then rising sharply above the canopy to reveal mountains at sunrise. One continuous shot, no cuts. Sound: wind rushing past.
Run 1 · BorderlineIt is shot from the drone's point of view, yet a small drone shows up near the ground.Run 2 · BorderlineIt is shot from the drone's point of view, yet a small drone shows up near the ground.Run 3 · BorderlineIt is shot from the drone's point of view, yet a drone flies through the frame.
Tests spoken line and mouth movement match. 98% usable across all models.
Prompt used for every model
Medium shot in a cozy cafe. A young woman with curly hair looks at the camera, smiles and says clearly in English: "I already ordered for you." Her lips move in sync with the words. Realistic, natural light. Sound: her voice and soft cafe chatter.
Run 1 · UsableRun 2 · UsableRun 3 · BorderlineThe line is right, but her hair is almost straight, not curly.
How we tested and scored
144 generations on November 2, 2026: 6 prompts, 3 runs each, on 8 models, all with the same settings and no retries. The hands shot first asked for dealing playing cards; card faces came out wrong so often that they got in the way of judging the hands, so on November 6 we swapped it for peeling an orange and ran all 8 models again. Only the orange clips count.
One question for every clip: could you drop it into a real project without generating it again? 1 = usable, nothing visibly broken and the shot does what it was asked. 0.5 = borderline, a small flaw you would only notice if you looked, or only part of the ask. 0 = not usable, something visibly broken (warped hands or faces, things appearing or vanishing) or the main ask missed.
Shot
Needed for a 1
Usually costs half a point
Full-body walk
Whole body in frame, walking toward the camera, legs and feet never merge, steady steps.
Feet slide for a moment, people in the background warp a little.
Hands doing a task
Hands look right (no extra, missing or merged fingers); the orange gets peeled and split into two halves.
Split with much of the peel still on, the fruit smears for a moment.
Face turning
Turns from profile to face the camera, stays the same person, smiles.
The turn stops short, features drift a little.
Readable text
The sign reads exactly OPEN 24 HOURS and stays that way.
Letters flicker in a few frames, ghost letters.
Fast camera move
One continuous shot through the trees that ends above the canopy with mountains in view.
Never clears the trees, trunks melt for a moment.
Dialogue and lip sync
Says the line (the transcript matches "I already ordered for you"); the same curly-haired woman, in a café.
A word added or dropped, mouth clearly out of sync.
Who scored
Claude, an AI model, scored every clip from frames taken three times a second, including the last frame, without knowing which model made it. A Whisper transcript checked the spoken line. Sound was not scored otherwise. Not counted against a clip: reflections and shadows that make physical sense, a face that looks like someone famous, and loose waves for "curly hair".
What this test can't tell you
One rater, and not a person. No one checked the scores against human ratings.
Three frames a second can miss a flicker shorter than that; lip sync was only checked roughly.
18 clips per model is enough to see big gaps, not to rank models a few points apart.
One setup: text to video, 5 seconds, 720p. Image to video, longer clips and other prompts may differ.
Questions
Which AI video model is cheapest per usable clip?
MiniMax H3: 20 credits for a 5-second clip and 92% of clips usable, so about 22 credits ($0.36) per usable clip in our test. Seedance 2.5 was the most expensive at $2.30.
Which model had the most usable clips?
Wan 3.0 scored 18 out of 18. All eight models scored between 81% and 100%, and with 18 clips per model one clip moves the rate by about 6 points, so most of those gaps are too small to rank models by.
Who scored the clips?
Claude, an AI model, against a fixed scoring sheet: it looked at every clip at three frames per second, and a Whisper transcript checked the spoken line. No people rated them. Every clip is on this page with its score, so you can check any of them.
Why do usable rates matter for cost?
You pay for every generation, including the ones you throw away. If one clip in four is unusable, each clip you keep costs a third more than the price per generation. In this test the usable rates were close, so the price per generation decided most of the cost.
Is this test sponsored?
No. VisionMorphic sells all eight models, and the credits on this page are our own prices. It is not affiliated with or endorsed by any of the model makers.
Run the prompts yourself
All eight models are in one generator with one set of credits. Copy a prompt from above, keep the length and resolution the same, and switch models.