AI video prompts tested: 8 hard shots on 8 models
Hands on a piano, a sign that has to read exactly right, water filling a glass, a crowd, a runner, someone eating, a mirror and a ball bouncing down stairs. These are the shots AI video has always struggled with. We gave 8 current models the same prompt for each and checked what they actually did. Every clip is real model output, one take each. Copy any prompt and try it yourself.
Scorecard: how many hard shots each model got right
| Model | Passed | Partly | Failed | Unclear | List price per clip | Output |
|---|---|---|---|---|---|---|
| Gemini Omni Flash | 7 | 1 | 0 | 0 | $0.50 | 1280×720, 5 s |
| Wan 3 | 6 | 1 | 0 | 1 | $0.50 | 1280×720, 5 s |
| Kling 4 | 5 | 3 | 0 | 0 | $0.64 | 1280×720, 5 s |
| Hailuo 3 | 5 | 2 | 0 | 1 | $0.50 | 1344×768, 5 s |
| Runway Gen-4.5 | 4 | 2 | 0 | 2 | $0.60 | 1280×720, 5 s |
| Seedance 2 | 4 | 2 | 1 | 1 | $1.50 | 1280×720, 5 s |
| Google Veo 3.1 | 4 | 0 | 0 | 4 | $1.20 | 1280×720, 6 s |
| Grok Imagine 1.5 | 3 | 2 | 0 | 3 | $0.80 | 1280×720, 5 s |
Out of 8 tests. "Unclear" means our two reviewers disagreed. Verdicts judge only the checks listed under each test, from still frames, not overall beauty. One take per model can't show consistency.
How we tested
We wrote one plain prompt per test and kept the camera still, so the hard thing is the only thing being tested. Every model got the same text and no starting image, 16:9, about 5 seconds. We ran each once through the Runway developer API on November 7, 2026 and kept every result, including the bad ones.
Clips are shown at each model's own output size: 1280×720 for most, 1344×768 for Hailuo 3. They are not upscaled. Veo 3.1 only offers 4, 6 or 8 seconds, so its clips are 6 seconds; all others are 5. The scorecard shows Runway's list price per clip at $0.01 per credit, no audio; our actual charges for many clips were higher than list, and the total above is what we really paid.
Each test has a short list of checks, like "five distinct fingers on each hand". Two reviewers judged every clip separately from 8 evenly spaced frames plus one full-size frame, without seeing each other's answers or which model made it. We publish a verdict only where both agreed; otherwise it says "Unclear". Still frames don't show smoothness, so treat the verdicts as a guide and watch the clip.
All clips are AI-generated test footage, not real filming. Generated with the Runway API.
Questions
Which AI video model handles hard shots best?
In this test, Gemini Omni Flash passed 7 of 8 hard shots and partly passed 1. That's one take per model per test, so treat it as a guide, not a ranking.
What can AI video still not do well?
The hardest test was running past the camera: 6 of 8 models did not clearly pass.
Can AI video models write readable text now?
In this test, 5 of 8 models clearly passed: the sign read exactly "OPEN LATE" and the rest of the scene matched the prompt. That's one short sign in one take, so longer or smaller text may still fail.
How much did it cost?
$63.32 for all 64 clips through Runway's API on November 7, 2026. Runway's list price per clip runs from $0.50 (Gemini Omni Flash) to $1.50 (Seedance 2); many of our real charges came out higher than list.
Are these clips upscaled or edited?
No. Each clip is exactly what the model returned, at its own output size, with no upscaling, trimming or colour changes.