Billing by how long the machine ran
Run-time billing charges for how long a model took to run rather than for the length of what it returned. It moves the variance of a slow queue or a hard prompt onto the buyer, and it needs a throughput figure to become a price for video. As of 2026-09-12.
| Channel | What the meter counts | Read from |
|---|---|---|
| Replicate | Most models by the time they take to run, with two video models per output second | The price page |
| Segmind | A dedicated endpoint by the hour, restated per second, plus a per-run price banded by steps | The pricing page |
| Channels metering output instead | The duration of the file returned, whatever it cost to produce | Their own rate pages, on the ledger |
Inclusion rule. Channels whose published meter counts machine time or invocations rather than seconds of delivered video, with one contrast row so the difference is visible. Order. Machine-time meters first, then the contrast.
1The buyer carries the variance
Under this meter a slow queue, a difficult prompt and an unlucky seed all cost money. Under output-second billing they cost the seller. Neither arrangement is unfair and they produce budgets of different shapes: one is a multiplication and one is an estimate with a tail.
For a production with a fixed delivery date the tail is the problem, because it arrives as an overrun nobody controls.
2Converting it needs a measurement, not a calculation
A machine-time figure becomes a video rate once somebody knows how many seconds of output a second of occupancy produces. That number depends on the model, the resolution and the settings, and no page here publishes it.
The consequence is that a machine-time figure cannot go on a ledger of rates. It can be recorded faithfully as what it is, which is what this register does.
3It usually means the platform is selling breadth
A catalogue of thousands of models cannot be priced model by model without enormous effort, and occupancy is the one meter that works for all of them at once. So run-time billing is a symptom of a wide catalogue rather than a pricing philosophy.
4Banding by effort is the same idea at a coarser grain
A per-run price banded by inference steps prices how hard the model was asked to try. Two buyers running the same model at different step counts pay different amounts for the same output length, which makes the figure legible and incomparable.
5What to measure before trusting it
The number that converts this meter is seconds of delivered output per second of occupancy, at the model and settings the work will actually use. It is a ten-minute measurement and it is the only way a machine-time figure becomes comparable with a rate.
Measure it on a batch rather than one call, and measure it at the busy hour. Both the queue and the prompt move the answer, and a figure taken from one idle test overstates the throughput a production will see.
6Pages these figures were read from
A definition entry: what the term means, and the figures vendors put in public about it, each figure read on 2026-09-12. Related: Conditional rate, Concurrency.
7The production behind the price
Price aside, what goes into the seconds behind “Run-time billing” is described in SceneMixer's guides:
Video is 16:9 landscape by default, with 9:16 vertical available; you set it once per project. Segments are typically 4 to 15 seconds, and some model tiers go to 30.
SceneMixer, AI Short Drama Generator — “What length and aspect ratio does it output?” (read 2026-10-13)
The shots add up to more than the selected video tier allows for one segment.
SceneMixer, Shot Seconds: Changing Timing, and the Limits — “What does 'over cap' mean on the segment card?” (read 2026-10-13)