Reliability, defined before it is measured
Most reliability pages are a victory lap. This one is a rulebook: how render time is counted, what counts as a failure, and the economic promise that makes the difference between the two your problem and ours.
Every engine on the roster is reachable from the same credit pool — no per-model plan, no cross-engine surcharge.
Short 720p clips from fast models land at the low end; 16-second takes and 4K renders sit at the top. The live quote before each run states what to expect.
When a generation fails, its credit charge reverses automatically — no support ticket, no waiting, no exceptions.
Download promptly after generation; both the clip and the source image are removed after a week.
The exact credit cost of the exact settings you picked is displayed before you commit — never discovered on the invoice.
One-time credit packs that never expire. Nothing renews, nothing to cancel.
What each model typically takes
Ranges reflect the model and workload, not queue luck: long-form and 4K renders compute longer by design. The quote shown before you commit is the credit cost — the clock below is what to plan around.
| Model | Typical render window | Known for |
|---|---|---|
| Grok Imagine 1.5 | ~2–4 min | Fastest social clips, native audio |
| Kling 2.6 | ~3–5 min | Cinematic motion, 720p |
| Seedance Mini | ~4–6 min | Budget everyday clips, audio included |
| Seedance 2.5 | ~5–8 min | Native 30-second clips |
| Veo 3.1 Fast | ~8–12 min | Maximum fidelity, 4K |
| Vidu Q4 | ~15–25 min | 16s long-form, up to 4K |
Why renders fail — and who pays
The single largest failure trigger is a prompt that trips upstream content filters — recognizable public figures, protected characters, flagged terms. Our pre-charge moderation screen intercepts most of these at submit time, before any credit moves. The remainder that slip through are rejected by the upstream model itself, logged as failures, and refunded on the spot.
The rest of the failure budget is provider errors and timeouts — again, never billed. Practical consequence: prompts that avoid real-person likenesses and protected IP clear the first-run gate far more often than average.
How we count
- Render time starts when the job is submitted (image upload already complete) and ends when the finished MP4 leaves our storage.
- A run is a failure when the upstream model returns no usable video — filter rejection, provider error, timeout, or a stuck job. Moderation blocks happen before any charge, so they are neither billed nor counted as failures.
- Every failed chargeable run auto-refunds. There is no manual claim and no threshold to meet.
- As this site accumulates volume, this page will publish the same discipline as numbers: 30-day delivery counts, success rates and per-model medians, computed exactly by the rules above.
First published November 5, 2026 · statistics section reserved until the first full 30-day window closes.
The guarantee behind the numbers
Generation is probabilistic; billing is not. No honest platform can promise every run succeeds — what can be promised, and what we promise, is the economics: money moves only when a video is delivered. Failed runs reverse themselves, credits never expire, and no subscription exists to forget about.
10 free credits toward your first video · no card required