RenderDrama

What a second of AI video costs, channel by channel

What Replicate meters on its price page

Replicate's default meter is time: most models are billed by how long they take to run. Two video models break that pattern with a figure per second of delivered output, which is the only shape on the page a buyer can budget from in advance. As of 2026-09-22.

Two meters on one catalogue, and who carries the varianceMost models here are billed by the time they take to run, which makes a slow queue, a hard prompt or an unlucky seed cost money. Two video models are billed per second of delivered output instead, which moves that variance onto the seller and makes a budget arithmetic rather than an estimate.Billed by run timeBilled per output secondWhat the meter countsHow long the machine was busyThe duration of the file returnedWho pays for a retryThe buyer, at full rateThe buyer, at full rateWho pays for a slow queueThe buyerThe sellerHow many modelsMost of the catalogueTwo video modelsBudget shapeAn estimate with a tailA multiplicationThe meter, not the model, decides whether a rate can be checked
Fig. 1 A run-time budget has a tail; an output-second budget does not. Only the second kind belongs in a price comparison without a large caveat attached.
The two meters Replicate prints, and which video models sit under each. Read from price page on 2026-09-22.
What is meteredAs the page prints it
Most modelsBilled by the time they take to run
One 480p video modelNine cents per second of output video
One 720p video modelTwenty-five cents per second of output video
What run-time billing pricesThe machine's occupancy, not the length of the result

Inclusion rule. Meters named on the price page. A model priced by run time and a model priced by output second are listed separately because the two cannot be reconciled without a benchmark. Order. The default meter first, then the models that depart from it.

1Where Replicate sits in this column

The unit each page counts inThe unit each of these channels puts its own figures against, Replicate marked. None of them is a plain price for one second of picture, which is why each gets an entry. Recorded 2026-09-22.The unit each page counts inWhat the page metersWhat the pagemetersfalfal — What the page meters: Seconds on most models, a whole clip on oneHeyGenHeyGen — What the page meters: Credits, as one currency for everythingMagnificMagnific — What the page meters: Credits against named durationsNovita AINovita AI — What the page meters: Seconds on one family, whole clips on twoReplicateReplicate — What the page meters: Run time, with two models by the secondSegmindSegmind — What the page meters: GPU time, plus a price per runSynthesiaSynthesia — What the page meters: Credits, converted into minutes
Fig. 2 The unit each of these channels puts its own figures against, Replicate marked. None of them is a plain price for one second of picture, which is why each gets an entry. Recorded 2026-09-22.
What the page meters, as each channel answering it has it, with Replicate marked. Recorded 2026-09-22.
ChannelWhat the page meters
falSeconds on most models, a whole clip on one
HeyGenCredits, as one currency for everything
MagnificCredits against named durations
Novita AISeconds on one family, whole clips on two
ReplicateRun time, with two models by the second
SegmindGPU time, plus a price per run
SynthesiaCredits, converted into minutes

Inclusion rule. Channels whose price page meters in something other than one second of delivered video. A channel printing money against a second is already comparable and belongs on the ledger instead. Order. Alphabetical by channel name.

2Run-time billing moves the risk onto the buyer

When the meter counts how long a machine was busy, a slow queue, a hard prompt or an unlucky seed all cost money. A per-second-of-output rate puts that variance on the seller: the buyer pays for six seconds of video whether it took twenty seconds or four minutes to make.

Neither is unfair, and they are not comparable. A budget written against a run-time model is an estimate with a tail; a budget written against an output-second model is arithmetic. Only the second kind belongs in a price comparison without a large caveat.

3Two video models with a published rate, and a wide gap between them

Nine cents a second at 480p against twenty-five at 720p is a step of nearly three times for one resolution increment, which is the steepest step recorded here on a single page. Both are the same model family; only the output size differs.

That matters because the lower figure is the one a comparison will quote. A production delivering at 720p should read the second row, and the second row is dearer than most of the register's headline numbers rather than cheaper.

4What the page does not carry

No maximum clip length appears on the page carrying the rate. Under run-time billing that omission compounds: a longer generation costs more because it is longer and costs more again because it occupies the machine longer, and neither multiplier has a published ceiling.

The workable approach is to treat the two per-second models as the priced part of the catalogue and everything else as a rate to be measured rather than read. That is a fair description of a platform whose product is breadth rather than a price list.

  • The meter used where no per-second video rate is printed
    Most models are billed by the time they take to runReplicate, prices / recorded 2026-09-22
  • wavespeedai/wan-2.1-i2v-480p
    $0.09 per second of output videoReplicate, prices / recorded 2026-09-22
  • wavespeedai/wan-2.1-i2v-720p
    $0.25 per second of output videoReplicate, prices / recorded 2026-09-22
  • Longest single generation
    Not stated on the vendor's price page (as of 2026-09-22)absent from the page carrying the rateReplicate, prices / recorded 2026-09-22

5Inside a generated second

Price aside, what goes into the seconds behind “What Replicate meters” is described in SceneMixer's guides:

The longest an episode may run: 240 seconds for Short Drama, 300 for TV Series, none for Film.

SceneMixer, Short Drama, TV Series or Film — “What is the episode cap?” (read 2026-10-13)

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)

6Sources

Read from the price page published at replicate.com on 2026-09-22. The rate this column sits beside is on Replicate; what the register accepts as published is on the method page. The same column on other channels: Novita AI, Segmind. Every column: what each channel publishes.