RenderDrama

What a second of AI video costs, channel by channel

The costs that never appear in a per-second rate

A per-second rate covers generating a second. It does not cover joining segments into an episode, which one channel prices separately at 1 credit per second of finished output, and it does not cover regenerating anything. As of 2026-09-12.

What each channel puts on its pricing pageFilled where the channel publishes something in that column, hollow where its pricing page is silent.What each channel puts on its pricing pageFigureWhat it coversCompositingCompositing — Figure: Priced separately on SceneMixerCompositing — What it covers: 1 credit per second of finished episodeRegenerationRegeneration — Figure: Not published by anyoneRegeneration — What it covers: No vendor states an expected attempt countClip ceilingClip ceiling — Figure: Published by one channelClip ceiling — What it covers: 30 seconds on Wan 3.0, 15 on MiniMax H3
Fig. 1 Filled where the channel publishes something in that column, hollow where its pricing page is silent.
What sits outside the per-second figure, and whether any channel publishes it. Recorded 2026-09-12.
ChannelFigureWhat it covers
CompositingPriced separately on SceneMixer1 credit per second of finished episode
RegenerationNot published by anyoneNo vendor states an expected attempt count
Clip ceilingPublished by one channel30 seconds on Wan 3.0, 15 on MiniMax H3

Inclusion rule. Cost components that apply after or around generation. A component no channel publishes is listed with that stated, because its absence is the finding. Order. Components a channel publishes first.

1Compositing is a separate line and usually a small one

One credit per second of finished episode is roughly a tenth of what the cheapest generation tier costs for the same second. On a fifteen-minute episode it is real money and it is not where the budget goes. It is worth recording because it is the only post-generation step any channel here prices at all.

The other channels do not mention joining because most of them are not selling an episode workflow; they sell clips and leave assembly to an editor. That is a product-shape difference rather than a pricing difference, and the row records it as such.

2The clip ceiling is a cost multiplier in disguise

A thirty-second ceiling and a fifteen-second ceiling produce the same minute at the same per-second rate, and they produce a different number of joins. Joins are where continuity breaks, and a break is regenerated, and regeneration is charged again at the full rate.

So the ceiling belongs in a cost record even though it is not a price. It is the published number with the clearest mechanical link to how many seconds a production ends up buying.

3Nobody publishes the number that decides the budget

How many attempts a usable shot takes is the multiplier on every figure in this ledger, and no channel publishes it. It would not be a stable number if they did: it depends on the material, the prompt and how fussy the production is.

Stating that plainly is more useful than estimating it. A ledger that invented a typical attempt count would be producing the single number readers most want and least able to support.

  • Compositing
    1 credit per second of finished episodebilled separately from generationSceneMixer, price list / recorded 2026-09-12
  • Longest single clip
    30 seconds on the Wan 3.0 and Seedance 2.5 tiers, 15 seconds on the Seedance 2.0 and MiniMax H3 tiersstated per tierSceneMixer, price list / recorded 2026-10-13

4Sources

Each figure is read from the vendor page linked in the ledger, checked 2026-09-12. How a figure qualifies is on method; what a rate leaves out is on limits and unspent credits and parallel generations. Nearby: Conditional prices, Paying for pixels, A minute, costed.