RenderDrama

What a second of AI video costs, channel by channel

Buying machine time is not buying video

Some pages price the machine rather than the result: an hourly figure for an endpoint, or a per-second figure for occupancy. Those numbers look like rates and are separated from one by a throughput figure the page does not give. As of 2026-09-12.

Four things a per-second figure can be countingA second of delivered video needs nothing further to be a rate. A second of machine occupancy needs a throughput figure, and the gap between the two readings can be a whole order of magnitude. A per-run figure banded by inference steps prices effort rather than either.Most usable at the topA second of deliveredvideoThe duration of the file returned. The only shape this registerputs in the ledger.A second of machine timeOccupancy. Needs how much output a second of occupancy produces,which nobody publishes.An hour of machine timeThe same meter at a coarser grain, usually for a dedicatedendpoint.One run, banded by stepsPrices how hard the model was asked to try. Legible, andcomparable with nothing.Under occupancy billing the buyer carries the queue, the prompt and the seed
Fig. 1 Three-tenths of a cent a second for hardware becomes anything from one and a half to nine cents a second of output, depending on a number no page publishes.
The two things a per-second figure can be attached to, and what each one leaves open. Recorded 2026-09-12.
What the second belongs toWhat the buyer pays forWhat is still needed
A second of delivered videoThe duration of the file returnedNothing: the figure is a rate
A second of machine occupancyHow long the hardware was busyHow many seconds of video that produces
A run, banded by effortOne invocation, priced by inference stepsBoth the output length and the step count a shot needs

Inclusion rule. The units a per-second or per-run figure can be attached to on the pages tracked here. A unit no page uses is not listed. Order. From the unit that needs no further figure to the unit that needs two.

1The gap between the two meanings is a whole order of magnitude

Suppose a page prices a machine at three-tenths of a cent a second. If a model returns six seconds of video for thirty seconds of occupancy, the real cost is about a cent and a half a second of output. If it needs three minutes, it is nine cents. Those are different purchases and the page cannot tell them apart.

So a machine-time figure is not a cheap rate. It is an input to a rate, and the multiplier has to be measured on the actual model at the actual settings before the number means anything.

2Who carries the variance is the real difference

Under output-second billing, a slow queue, a hard prompt and an unlucky seed cost the seller. Under machine-time billing they cost the buyer. Neither is unfair and they produce budgets of different shapes: one is a multiplication, the other is an estimate with a tail.

For a production with a delivery date, the tail is the problem. A month of generation priced by occupancy can come in twenty per cent over for reasons no one controls, and the same month priced per output second cannot.

3Pricing by effort is a third shape

A per-run figure banded by inference steps prices how hard the buyer asked the model to try. It is legible and it is not comparable with anything: two buyers running the same model at different step counts pay different amounts for the same length of output.

The workable approach with all three shapes is to treat only the output-second figures as prices and everything else as something to benchmark. A register can record the others faithfully; it cannot convert them.

Written to explain, not to record. No sentence here is attributed to a seller, and none of it should be quoted as a fact about one. The sourced material is on the ledger. Related: Limits and deadlines, Bands and grids.

4What the seconds buy

Price aside, what goes into the seconds behind “Buying machine time” is described in SceneMixer's guides:

Every character gets an AI-generated full-body reference sheet at analysis time. Face and body crops from that sheet are injected into every video segment that features the character, so the same face, hair, and signature look persist across all episodes.

SceneMixer, Turn a Novel into a Video Series with AI — “How does SceneMixer keep characters consistent across episodes?” (read 2026-10-13)

Every segment starts with a short header that sets the stage before the first shot: Location (a scene chip), Time, Weather, the Opening frame (what the first frame shows), Set dressing (what is in the room, and any background people) and Props (the objects the segment uses).

SceneMixer, The Segment Header, Field by Field (read 2026-10-13)