Guides

Price per Second Is Not Cost per Usable Clip: A Practical Test for Cheap Video APIs

3 November 2026 · RenderBob team

A video API at half a cent a second can still cost more per finished shot than a dearer one, if more of its clips are thrown away. Measure cost per usable clip on your own images.

A video model that costs half a cent a second can still cost more per finished shot than one that costs ten times as much, if more of its clips get thrown away. Cost per usable clip is the comparison that holds.

The rate card leaves out the keep rate

With Mischief 1 listed at $0.005 per second, it is tempting to compare it with other services on that number alone. An AI Idelist comparison, reviewed on 11 October, shows how crowded the cheap end of Wan 2.2 already is. On that page, CFLabs lists a Turbo variant at $0.005 per second at 480p, and Eternal AI lists a Lightning image-to-video model at the same 480p rate. fal's standard Wan 2.2 image-to-video listing is $0.04 per second at 480p and $0.08 at 720p. Those prices cover different variants, tiers, and billing rules. The page says its totals exclude retries, storage, transfer, and taxes, and it claims no quality or reliability test. Its "cheapest" labels are the page's own ranking.

The formula

Cost per usable clip is the price of one attempt divided by the share of attempts you would keep. The numbers below are made-up examples, not measurements.

Illustrative examples, not a named test. A clip costing $0.025 where one attempt in four is usable costs $0.10 per usable clip. A clip costing $0.20 where three attempts in four are usable costs about $0.27. If the cheap model's keep rate fell to one in ten, the first example would be $0.25 per usable clip.

The ranking can change on the keep rate alone, which is why it has to be measured on your own material.

What else changes the answer

• Resolution and length. A price listed at 480p, or for one clip length, may not match what you will ship.
• Billing rules. Some listings bill per second with a five-second minimum. Others bill per video.
• Image fidelity. For image-to-video, how closely the clip keeps the start frame decides how much cleanup follows.
• Storage. The AI Idelist page notes that one provider's generated video links expire after about 24 hours, so outputs have to be copied elsewhere.
• Licence and resale terms. The same page notes that one provider restricts reselling without consent.
• Queue time. A fast model that is often busy is not fast.

A test you can run in an afternoon

• Pick ten start images that represent your real work, including hard ones.
• Write one motion prompt per image and keep it fixed across services.
• Run each service at the setting you will ship, for example 720p and five seconds.
• Score every clip as keep or discard, using a rule written before you look.
• Record price, queue time, and total attempts, including failures.
• Compute cost per usable clip and time per usable clip, then repeat with a second set of images.

What to watch

Independent comparisons that publish keep rates as well as prices, billing changes as vendors add faster variants, and whether anyone starts guaranteeing a cost per usable clip rather than a rate per second.

RenderBob 3 and this story

RenderBob 3 keeps the model a per-clip choice and keeps each clip's references and prompt log with the studio, so a test like this can be repeated and the cheapest usable option chosen shot by shot.