How per-second billing actually works

A per-second rate prices one second of output that the service considers delivered. It does not price the attempts you discard, the joins between clips, or the time spent waiting. Those are real costs and none of them is in the number. As of 2026-09-12.

Inclusion rule. The three shapes a published price takes across the channels tracked here, counted on the ledger. A monthly allowance counts none there because the ledger never converts one into a per-second figure. Order. Ascending by how much arithmetic the buyer is left with.

1The unit is a second of output, not a second of work

When a service says nine cents a second, the meter is counting the duration of the file you get back. A five-second clip costs five times the rate whether the model took twenty seconds or four minutes to produce it. That is good for the buyer in one specific way: your cost is predictable from the length you asked for, and it does not move because the platform was busy.

It also means the rate tells you nothing about throughput. Two services at the same rate can differ by a factor of five in how long you wait, and waiting is the cost that shows up in a schedule rather than an invoice. If a production has a delivery date, the queue matters as much as the price, and almost nobody publishes queue behaviour.

2Retries are the multiplier nobody prints

Generated video is iterative. A shot gets regenerated because a hand came out wrong, because the character drifted, because the camera move was not what the prompt meant, or because the director changed their mind. Each of those attempts is a full-price second.

There is no published attempt ratio anywhere in this field, and there could not be an honest one: it depends on the material, the prompt, the model and how fussy the production is. What a team can do is measure their own. Keep a count of generated seconds against delivered seconds for one episode and you have a multiplier that is true for your work, which is worth more than any industry figure.

A useful early estimate is to budget the delivered duration, then multiply. Teams that have never measured are usually surprised by how large their own number is, and the surprise arrives at the end of the project rather than the start.

3Resolution is a price lever and usually the wrong one to pull

Nearly every service charges more per second at higher resolution, often one and a half to five times as much at the top of the ladder as at the bottom. For vertical drama watched on phones, the top of the ladder is frequently paying for pixels the delivery path throws away: platforms re-encode, and a phone screen has a ceiling.

The case for generating high is reframing. If a shot might be cropped, punched in, or stabilised in the edit, extra pixels are headroom you cannot add later. The case against is everything else. Deciding this per project rather than per company is the cheapest optimisation available in this whole workflow.

4Credits are a second currency with a floating exchange rate

Some services price in credits and sell credits in packs, where a bigger pack lowers the price per credit. The same generation therefore has two correct dollar figures depending on what was bought, sometimes differing by nearly half.

Two consequences follow. First, any comparison between a credit-priced service and a currency-priced one has to state which end of the pack range it used, and a comparison that does not is quietly choosing the flattering end. Second, per-project accounting becomes harder: the cost of this episode depends on a pack bought weeks ago at a rate nobody recorded. Writing down the pack rate on the day of purchase takes a minute and saves an argument later.

5The clip ceiling is a cost figure wearing a technical disguise

Every service has a maximum length for one generation. That number does not appear to be about money, and it is: a sixty-second piece assembled from fifteen-second clips has three joins, and a join is where continuity breaks. Breaks get regenerated. Regeneration is charged.

So when two services quote the same rate and one allows twice the clip length, they are not the same price for long-form work. This is the single most under-weighted number when people compare generation costs, and it is often not printed on the pricing page at all.

6What to write down before you commit a budget

Four things, none of which requires a vendor to tell you anything: the delivered duration you actually need, your own measured attempt ratio, the resolution the delivery path can use, and the number of joins your clip ceiling forces. Multiply those against a published rate and you have an estimate you can defend.

Everything else in a pricing comparison is noise until those four are settled. A rate that is twenty per cent cheaper cannot rescue a plan whose attempt ratio is three, and a plan with a measured ratio of one point two is comfortable at almost any published rate in this market.

7Where the sourced entries live

This page is about how the pricing works. What any particular channel charges is a separate matter, kept in the ledger with the page each figure came from and the date it was read.

• SceneMixer — prints the model and version beside each per-second rate
• Kling AI — publishes developer prices by resolution
• LTX Studio — publishes an API rate as a band rather than a figure
• Google — publishes Veo rates on its API pricing page

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: Budgeting a series, Costing a season.