Skip to content

Seedance 2.0 beat Gemini Omni 1.2 Flash in 5 video tests

Illustration of a woman in a leather jacket stamping WIN on a film strip while another strip burns in a graffiti editing room Models

Seedance 2.0 beat Gemini Omni 1.2 Flash 4 to 2.5 in 5 tests of things video models usually get wrong. Omni was faster and cheaper: 47 seconds and $0.28 per 4-second clip, against about 2m 50s and $0.61 for Seedance. But Omni failed two tests and Seedance failed one.

Claude Opus 5.5 tests Gemini Omni 1.2 Flash against Seedance 2.0 on 5 scenes video models usually break

The setup

Google released Gemini Omni 1.2 Flash on Nov 5, 2026. We asked Claude Opus 5.5 in Claude Code to test it against Seedance 2.0 from ByteDance.

  • Models: Gemini Omni 1.2 Flash and Seedance 2.0, both at 720p
  • One prompt per scene, the same for both models
  • 5 scenes, 4 seconds each in the video
  • Test date: Nov 6, 2026

Opus picked what video models usually break: hidden objects staying the same, solid walls, body weight, hits that connect and motion that keeps its direction. It wrote the prompts, ran the Omni clips and scored both models from the frames. We checked every score by eye and changed two of them.

The 5 tests

Each scene checks one thing that breaks often in AI video.

  1. Hidden objects: a golden retriever runs behind a thick oak and must come out the other side
  2. Solid walls: a steel ball rolls through a wooden maze and must never pass through a wall
  3. Body weight: a parkour runner jumps a 3-meter gap between roofs, lands and rolls
  4. Hits that connect: two fighters trade punches and a spinning kick in a warehouse
  5. Direction: a longboarder carves down a mountain road while a drone follows
Gemini Omni 1.2 Flash and Seedance 2.0 side by side: a golden retriever running past a big oak tree in an autumn park
Hidden objects: both models keep the same dog

The results

A pass is 1 point, a partial pass is 0.5 and a fail is 0.

TestGemini Omni 1.2 FlashSeedance 2.0
Dog behind the treePassPass
Ball in the mazeFail: the ball doubles and teleportsFail: the ball passes through a wall and rolls by itself
Rooftop gap jumpPassPass
Warehouse fightPartial: the hit reaction is offPass
Downhill longboardFail: the rider turns and rolls back uphillPass
Score2.54

Neither model could keep a ball inside a maze. In Seedance's clip, nobody tilts the board, yet the ball keeps rolling and goes through a wall.

Two AI video clips of a steel ball in a wooden maze toy, from Gemini Omni 1.2 Flash and Seedance 2.0, both marked as failed
Solid walls: both models fail
Gemini Omni 1.2 Flash and Seedance 2.0 clips of a parkour runner in a hoodie jumping between city rooftops at sunset
Body weight: both pass
Gemini Omni 1.2 Flash and Seedance 2.0 clips of two fighters in a dusty warehouse, Omni marked as a partial pass
Hits that connect: Omni's hit reaction is off
Gemini Omni 1.2 Flash and Seedance 2.0 clips of a longboarder in a red jacket on a mountain road, Omni marked as failed
Direction: Omni's rider rolls back uphill

The numbers

Gemini Omni 1.2 FlashSeedance 2.0
Time per 4-second clip47s2m 50s
Cost per 4-second clip$0.28$0.61
Generations85
Time, all generations7m 20s18m
Cost, all generations$3.06$3.66

Omni needed 3 attempts at the 8-second downhill clip and 2 at the parkour jump. Even with those retries, it spent less money and less than half the time. Its third downhill attempt used a prompt we edited to force the rider forward, so we did not use that clip in the video.

Animated verdict screen with Claude Opus 5.5 and confetti naming Seedance 2.0 the winner over Gemini Omni 1.2 Flash
The verdict: Seedance 2.0 wins 4 to 2.5

Our take

Seedance 2.0 is the safer pick when the shot has to work on the first try. Gemini Omni 1.2 Flash is about 3.6x faster and half the price per clip, but be ready to redo the fails. It matched Seedance on the dog, the jump and most of the fight. Both still break basic physics and a ball in a maze was enough to show it.

Sources

Portrait of Yana Lialuke, author at thehype.news

Yana Lialuke

Content Creator at thehype. Yana tests AI models hands-on across image generation, 3D, and LLMs. Same prompts, real costs. An AI marketer turned creator, she builds with agents, from 3D games to Three.js scenes, sharing prompts and failures.

More articles by Yana Lialuke →
On air · radio.thehype.news ↗ AI news radio — 24/7