The Real Price of One Usable AI Shot Is Not on the Rate Card
Per-second AI video prices ignore rerolls and operator time. A worked 8-shot example and a quoting formula that stops you underpricing.
Cover illustration generated with AI for Fix It In Post.
An AI video rate card tells you the cost of one generation. A client quote needs the cost of one shot you can actually use. These are different numbers, often by a factor of five to twenty-five, and the gap is where small studios lose money on AI work.
Our position is simple: never quote AI shots per second of output. Quote per usable shot, with the keeper ratio and the operator hours written into the estimate. Below is the method, a worked example for a 30-second, eight-shot spot on three platforms, and the multipliers we would use.
What the rate cards say
Start with the list prices, because everything else multiplies them.
- Kling 3.0 bills in credits: 6 per second at 720p without audio, 12 per second at 1080p with audio, 30 per second at native 4K. At the standard rate of 66 credits per dollar, that is roughly 18 cents a second for 1080p with sound, according to CloudZero's breakdown of Kling's credit guide. We could not load Kling's own billing page to check this independently.
- Veo on Google Cloud: Google's Vertex AI pricing page lists Veo 3 at $0.40 per second with audio and $0.20 without, and Veo 3 Fast at $0.15 with audio. For Veo 3.1, an aggregator analysis reports about $0.40 per second at Quality, $0.10 at Fast and $0.05 at Lite. We could not find Veo 3.1 rates on Google's own pricing pages, so treat those three figures as reported, not confirmed.
- Google Flow: the AI Pro plan costs $19.99 a month for 1,000 credits. A Fast clip costs about 20 credits and a Quality clip about 100, per CostGoat's Flow pricing tracker (an aggregator; Google does not publish a per-clip table we could check).
The number the rate card leaves out
The keeper ratio is how many generations you burn for each shot that survives review. Published estimates vary a lot:
- CloudZero assumes a 70% first-take success rate, which works out to a 1.4x multiplier. That turns a 120-credit, 10-second 1080p Kling clip into about 168 credits, or $2.55 per finished clip.
- The ocdevel analysis suggests budgeting four generations per usable shot when the composition and characters are specific. On that basis it puts frontier clips with audio at $3–13 per usable shot, and open-weight models at under $1.
- PJ Accetturo's widely shared Kalshi spot for the 2025 NBA Finals took 300–400 Veo 3 generations to get about 15 usable clips, roughly $2,000 in credits over two to three days. That is 20 to 27 takes per keeper, from a director with more than 15 years' experience.
The 1.4x figure suits an unspecific B-roll shot where almost anything on-brief will do. The Accetturo figure is what a commercial with recurring characters, specific action and comic timing really costs. Most client work falls somewhere between the two, and closer to the high end than people expect.
Worked example: 30 seconds, eight shots
Here is the brief: eight shots cut to 30 seconds, each about four seconds on screen. We generate an 8-second clip per take on Veo (the usual length) and a 10-second clip on Kling, because you need handles. The table shows generation cost only, across four keeper ratios.
| Platform (per take) | 1 take/shot | 1.4x | 4x | 25x |
|---|---|---|---|---|
| Veo Quality via API, 8 s at $0.40/s ($3.20) | $25.60 | $35.84 | $102.40 | $640 |
| Kling 3.0, 10 s 1080p + audio, 120 credits ($1.82) | $14.55 | $20.36 | $58.18 | $363.64 |
| Flow Quality, 100 credits at Pro plan rate ($2.00) | $16.00 | $22.40 | $64.00 | $400 |
Two things stand out. First, even the worst case is a few hundred dollars, so generation fees are rarely what sinks the job. Second, Flow's subscription cap matters more than its price. Eight shots at four Quality takes each is 3,200 credits, more than three months of the AI Pro allowance. At 25 takes a shot you need 20,000 credits, which means a higher tier or a second account. Check the cap against the schedule before you promise a delivery date.
The real cost is the operator
Every take has to be prompted, waited for, watched and judged. Assume six minutes per take, which is our working figure and not a published one, and an operator at $75 an hour:
- At 4 takes a shot: 32 takes, 3.2 hours, about $240 of labour against $58–102 of generation.
- At 25 takes a shot: 200 takes, 20 hours, about $1,500 of labour against $364–640 of generation.
Twenty hours is two and a half working days, which is roughly what Accetturo reported. Then add the conform: upscaling, grading AI clips to sit together, fixing hands and text in post, sound. Budget at least half a day for that on a 30-second spot.
The quoting formula
Use this per shot, then sum:
- Take cost = rate per second × generated clip length (not on-screen length).
- Keeper multiplier: 1.5x for generic atmospherics, 4x for specific compositions, 10x for anything with a recurring character or product, 25x for hero shots that need performance or comic timing. The 10x tier is our own interpolation; the others come from the sources above.
- Generation line = take cost × multiplier.
- Operator line = minutes per take × multiplier ÷ 60 × hourly rate.
- Conform line: a fixed block per spot, plus extra for any shot that needs paint-out or upscaling.
- Contingency: 20% on the generation line, because prices change monthly and models are retired without much warning.
Put the multiplier on the estimate where the client can see it. That explains why a hero shot costs more than a sky, and it gives you room to talk when the client wants "just one more version". If a client is comparing your AI quote with someone's per-second rate card, show them this table.
The takeaway
The per-second price is the smallest number on an AI job and the easiest to quote, which is why people underprice. Quote per usable shot, with an explicit keeper multiplier and operator hours, and recheck rate cards every quarter. If your quote is mostly generation fees, you have almost certainly left out the labour.
