What one second of AI video actually costs
Measured on real GPUs, priced from real invoices. Every figure below comes from a job we ran and paid for. Where we don't have a number, the cell is empty rather than estimated.
Last updated 2026-10-14.
Why this page exists
We went looking for an honest comparison and couldn't find one. What exists is vendor marketing, blog posts quoting each other, and research tables reporting time per diffusion step rather than wall clock. None of it tells you what you will actually be billed.
Three different numbers get called “speed”
- Warm compute — GPU seconds per second of video, weights already loaded. This is what vendors quote.
- Start to download — what you actually wait, including allocation, boot and model download.
- Dollars per video-second — money debited divided by video delivered. Only this one is a cost.
A model that generates in 9 seconds behind a 147-second boot does not cost what the 9 seconds suggest. For one clip, the boot is the cost.
The finding that matters most
The dearest and cheapest priced rows here are the same model, on the same GPU, on the same day — and differ 2.5× in cost per video-second ($0.2376 against $0.0935).
The only material difference is how much resolution the reference image is given — a configuration choice most comparisons never mention. If you are deciding where to spend attention: tune your configuration before you shop for a model.
Measured runs
| Model | GPU | Resolution | Output | Wall clock | Warm s/video-s | $ / video-second |
|---|---|---|---|---|---|---|
MiniMax H3 timeline full reference | B200 secure | 756x1344 | 16.2 s | 26 min 08 s | — | $0.2376 |
MiniMax H3 timeline 384x736 reference | B200 secure | 736x1394 | 18.5 s | 11 min 51 s | — | $0.0935 |
LTX-Video 0.9.8 2B 8 steps, euler | A6000 | 768x512 | 3.9 s | — | 2.2 s | $0.0004 warm only |
MiniMax H3 timeline (full reference): Wall clock includes allocation, boot and model download. Warm compute was not isolated, so it is not reported.
MiniMax H3 timeline (384x736 reference): Same model, same GPU, same day as the row above. 2.5x cheaper per video-second on reference budget alone.
LTX-Video 0.9.8 2B (8 steps, euler): Replicated on a second pod at $0.045. The warm figure excludes a 147-second boot; for one short clip the boot is most of the cost. Fast, but adherence measured 2 of 5: misses exact object counts, camera direction and on-screen text, identically across seeds.
Figures we have not re-verified
Recorded in our notes when we measured them, but without a transaction receipt we have re-checked. Reported here because leaving them out would be less honest than labelling them.
- MiniMax H3 — 8.7 s of warm compute per video-second on A6000 class, measured 2026-08-17. Recorded in our notes at the time. Transaction receipt not re-verified, so treat as indicative until re-run.
- Wan Animate — 35–60 s of warm compute per video-second on A6000 class, measured 2026-08-17. Recorded in our notes at the time. The range is the observed spread, not one run.
What this page cannot tell you
- Which model is better. The cheapest row here misses object counts, camera direction and on-screen text. Cheap and usable are different axes.
- Costs on hardware we have not run. No row is extrapolated.
- Today's provider rates. Each row is priced at the date shown, and rates move.
- A controlled benchmark. Rows differ in GPU, resolution and pipeline. Comparing two rows that differ in more than one dimension proves nothing.
Want a model or GPU that isn't here?
Every row costs a real GPU run, so we add one when it would change somebody's decision. In the meantime, the free workflow checker tells you which GPU your own workflow needs before you deploy it — no signup.
Check a workflow