Same prompt, different AI video models: 4 side-by-side tests

Update, October 1, 2026: PixVerse v6 was added to tests 1 and 2.

Model pages tell you what a model can do. These four tests show what it did with the same input. Each clip below was generated on Cicadas and is in the test log with its request ID; we show every run we made for these tests, including the ones that missed the prompt.

Two things to keep in mind. First, three to four clips per test is far too few to rank models; treat each result as one data point. Second, lengths differ where a model has no shorter option: Seedance ran 4 seconds; Kling, Vidu Q3, PixVerse v6 and FLUX 3 ran 5.

Test 1: one prompt, four models (paper city)

Prompt: a handcrafted paper city at blue hour, a small red tram gliding past, window lights turning on one by one, gentle paper-wheel sounds. Text to video, 16:9, sound on.

| Model | Settings | Credits today | Provider cost | Request | |---|---|---|---|---| | Seedance 2.5 | 4 s, 480p | 221 | ≈$0.88 | 01a0c932-aeb2… | | Seedance 2.0 | 4 s, 480p | 136 | ≈$0.54 | 01a0f169-0be5… | | Vidu Q3 | 5 s, 540p | 88 | ≈$0.35 | 01a0f22c-a4af… | | PixVerse v6 | 5 s, 540p | 59 | ≈$0.23 | 01a0f4d9-e5a1… |

All four got the look right. The difference is timing and sound. Seedance 2.5 lit the windows while the tram crossed, with rattling and rustling in the audio. Seedance 2.0 did the same, but our listening model heard a steady water-like sound rather than wheels. Vidu Q3 sent the tram across and out of frame before any window lit, with a soft rustle. PixVerse v6, the cheapest of the four, sent the tram across the whole frame, but every window was already lit in the first frame; the sound was again a soft rustle.

Test 2: one kitten photo, four models

Kling 3.0 Pro, Vidu Q3 and PixVerse v6 ran the exact same prompt (one blink and a small head tilt, paws planted); the Seedance prompt was written for Seedance. All silent.

| Model | Settings | Credits today | Provider cost | Request | |---|---|---|---|---| | Kling 3.0 Pro | 5 s | 140 | $0.56 | 01a081a1-77fb… | | Seedance 2.5 | 4 s, 480p | 221 | ≈$0.88 | 01a0ce04-4477… | | Vidu Q3 | 5 s, 720p | 193 | ≈$0.77 | 01a0f22e-224d… | | PixVerse v6 | 5 s, 720p | 57 | ≈$0.23 | 01a0f4da-4465… |

Seedance stayed closest to "one small movement", with softer fur. Vidu Q3 moved more than asked, turning the head from side to side. PixVerse v6 kept the paws planted but turned the whole head to the window, and the framing crept closer although the prompt asked for a locked camera. Kling, Vidu Q3 and PixVerse v6 all kept the photo's 3:4 shape even though the request said 16:9 or 9:16, which is how image-to-video works on these models.

Test 3: one ring illustration, three models

Kling 3.0 Pro and FLUX 3 ran the exact same prompt (the ring stays still while a narrow light sweeps across it); the Seedance prompt was written for Seedance. All silent.

| Model | Settings | Credits today | Provider cost | Request | |---|---|---|---|---| | Kling 3.0 Pro | 5 s | 140 | $0.56 | 01a0819f-3cce… | | Seedance 2.5 | 4 s, 720p | 473 | ≈$1.89 | 01a0cd35-00d0… | | FLUX 3 | 5 s, 720p | 213 | ≈$0.85 | 01a0f22c-8811… |

All three kept the ring's shape. FLUX 3 kept it so still that the light barely moved in five seconds; the clip reads almost as a photograph. For a product close-up that may be what you want; for a visible light sweep it is not.

Test 4: one coastal image, one prompt, two models

The same image and the same long prompt: a graphite coupe driving away along a wet cliff road, a restrained descending camera, no new vehicles. Silent.

| Model | Settings | Credits today | Provider cost | Request | |---|---|---|---|---| | Seedance 2.5 | 4 s, 720p | 473 | ≈$1.89 | 01a0ce7b-f044… | | Kling 3.0 Pro | 5 s | 140 | ≈$0.56 | 01a0f16e-b3bf… |

Seedance followed the shot as written, with foliage and the guardrail moving faster than the distant city. Kling made a bigger camera move than asked and added two things the prompt ruled out: a second car far up the road and a large wave bursting against the cliff. It also cost a third as much.

What these tests do and don't show

• The same prompt can come back with different timing, motion and sound. Check the whole clip, and listen, before you publish.
• Price per clip differs by more than five times for similar lengths, so try a cheap model first when the look matters more than exact motion.
• Nothing here ranks the models. For what each one is good at, see Model alternatives; for every price, see Pricing.