Controlled model test · July 2026

One prompt. Four submissions. Three completed renders.

The same product-shot brief went to four text-to-video models. We publish every outcome, including the request that never reached inference.

July 13, 2026

4Submitted
3Completed
1Did not start
0Retries
Same prompt for every model

Verbatim test prompt

One continuous widescreen 16:9 product-commercial shot. A coral-orange sneaker sits on a matte black pedestal and remains the exact same shape and color throughout. The camera dollies slowly from left to right while exactly three clear glass spheres roll behind the pedestal. Warm orange rim light, realistic fabric and rubber texture, physically believable motion, premium studio photography. No people, no text, no logos, no cuts, no camera shake.
Completed outputs

Four-frame review

Download machine-readable data
1280x720 · 4s

Veo 3.1 Lite

Completed
End-to-end time
34.3s
Provider inference
27.2s
Provider cost
$0.12
Catalog credits
24
Shoe continuityPass
Exactly three spheresPartial
Spheres behind pedestalPartial
No visible textPass
1280x720 · 5.04s

Grok Imagine

Completed
End-to-end time
56.7s
Provider inference
51.4s
Provider cost
$0.35
Catalog credits
60
Shoe continuityPass
Exactly three spheresMiss
Spheres behind pedestalPass
No visible textPass
1280x720 · 5.04s

Seedance 2.0 Mini

Completed
End-to-end time
120.8s
Provider inference
113.8s
Provider cost
$0.61
Catalog credits
85
Shoe continuityPass
Exactly three spheresPartial
Spheres behind pedestalMiss
No visible textPass
Did not start

Hailuo 02

The provider returned “payment required” before creating a prediction. No model inference ran, no output exists and the submission remains in the report.

90 Catalog credits
01 · Protocol

How the checks were made

Each model received one submission with the exact prompt above, a 16:9 target, no audio and zero retries. One reviewer inspected four evenly distributed frames from every completed output. The original outputs are embedded so readers can inspect them directly.

02 · What this test does not prove

What this test does not prove

This is one prompt, one run and one reviewer—not a universal model ranking. Frame checks can evaluate visible continuity and composition, but they do not fully score motion quality. Provider time and cost can change.

Controlled AI video model benchmark · July 2026 — Reelipal Research