How we review AI video models.
Source review comes first. Identical hands-on test prompts follow when access is available. A page never implies original output until the exported video and its settings are published.
Source review
We collect official documentation, model cards, system notes, release posts, and current pricing pages. We write down the capability, restriction, or commercial term exactly enough for a reader to inspect the primary source.
Matched prompt tests
We run the same prompts and settings across models wherever access allows. This isolates model behavior rather than rewarding a prompt that happens to fit one interface better.
Transparent updates
We publish methods, prompts, settings, and dates. When models or pricing change, we update the records and note what changed instead of letting an old result silently stand.
What a completed test contains
Every item is visible on the comparison page
What we hold constant
The identical English prompt is submitted to every model in a test round. Image-to-video tests use the same source image, with its license, dimensions, and preprocessing disclosed.
We match duration, aspect ratio, resolution, audio setting, seed, and prompt-expansion setting whenever those controls exist in both products. When a setting cannot be matched, the comparison states the difference rather than treating a higher tier as an apples-to-apples win.
Evidence rules
Each page shows the full prompt, generation date, plan, settings, video duration, poster image, and an unedited still from every compared output. Product claims and prices link to first-party sources. Vendor benchmarks are described as vendor claims; they never substitute for our own output evidence.
What we score
- Motion consistency. Stable body mechanics, trajectories, camera movement, occlusion, and scene contact from first frame to last.
- Character consistency. The supplied or described subject keeps recognizable face, wardrobe, proportions, and identity through motion and cuts.
- Visual detail. Clean textures, anatomy, lighting, depth, and artifact handling at the exported resolution.
- Text rendering. Legible, correctly spelled, stable lettering in a dedicated sign-and-product-label test.
- Style control. The result follows the requested visual language without replacing it with the model's default look.
- Speed and reliability. Median submit-to-ready time, queue errors, and failed-job rate over the stated sample.
- Value. The current price for a usable clip, including resolution, duration, reference inputs, and likely retry cost.
- Ecosystem and API. Documented public API access, controls, asset workflows, rates, integrations, and commercial-use terms.
Scoring and revisions
10-point scale when assets exist
Scores use a 10-point scale. Two editors score motion and character consistency independently; a disagreement greater than one point is adjudicated from the exported file. We recheck subscriptions, credit rates, model versions, and API availability whenever a page is updated.