What an AI short actually costs to finish

The number nobody quotes is attempts per keeper. Everything else follows from it.

Ash Koosha made "Dream of Violets" for about $2,000 (roughly £1,477) and took it to Tribeca. His follow-up, the 135 minute "Odysseus: The Fall", is budgeted in the mid five figures, and he rebuilt the pipeline around Kling once OpenAI began winding Sora down. In the same season Christopher Nolan's "The Odyssey" reached theaters on 17 July carrying a budget reported at around $250 million. Both films are adaptations of the same poem. The gap between them is the entire argument about what generated filmmaking costs, and most of that argument is being conducted with the wrong numbers.

The wrong numbers are list prices. Veo 3.1 Standard is $0.40 per second with audio at 1080p, which makes an eight second shot $3.20 (£2.36). Seedance 2.5 bills $0.2311 per second at 720p and $0.5686 at 1080p, so a five second 16:9 clip lands at $1.16 or $2.84. Runway Gen-4.5 charges 60 credits for five seconds, about $0.12 per second on a $28 Pro plan. Kling Video 3.0 spends eight credits per second at 1080p without audio and twelve with native audio, and its cheapest commercial tier is $6.99 a month for 660 credits.

Read those figures straight and a two minute film costs less than lunch. That is not what anyone actually pays, because a generated shot is not a shot until you accept it.

What the yield does to the price

Practitioners running production volume report roughly three generations per usable shot, and then throw most of the usable ones away in the edit. One documented three minute animated episode used 164 Seedance clips to fill 41 slots in the final cut, a selection rate near 25 percent. A 90 second horror short reportedly consumed about 400 generations. Chinese outlet 36kr reported on 1 April that attempts per keeper had moved from two or three up to seven or eight during the spring capacity crunch, as queues lengthened and models were served at lower effective quality.

Stack the two effects and the multiplier is not 3x, it is closer to 10x. Three attempts to get a shot you would show anyone, then four shots kept out of every ten you would show. A 15 second generation typically yields four to seven seconds you would cut with.

So we ran our own test rather than repeat someone else's ratio.

Our own run, priced to the second

One prompt, four models, same night: a woman in a wet trench coat standing beside a red London double decker at night, neon reflections on the road, handheld. We asked each model for a single take at its own native settings and paid list price.

• Veo 3.1 Fast with audio, 6 seconds, 1080p: $0.72 (£0.53), delivered in 63 seconds.
• Kling Video 3.0 Standard, 5 seconds, 1080p without audio: 40 credits, about $0.35 (£0.26) on a $25.99 Pro plan, delivered in 88 seconds.
• Seedance 2.5 at 720p, 6 seconds: $1.39 (£1.03), delivered in 458 seconds.
• Seedance 2.5 at 480p, 6 seconds: $0.62 (£0.46), delivered in 506 seconds.

Total spend for 23 generated seconds: $3.08, or £2.27 at 0.7384 to the dollar.

Then we graded them against the brief, which is the part the price lists cannot do for you. Veo returned the shot we asked for, correct bus, correct framing, correct weather, with the destination blind rendered as illegible smear. Kling returned a technically clean and well lit plate with a white and blue single decker in it, which is a bus but not the bus. Both Seedance takes drifted: the 720p attempt put the subject in a blue coat with no bus in frame at the two second mark, and the 480p attempt found the red bus but softened the whole image to a resolution we would not cut into a 1080p timeline.

One usable take out of four attempts, on a prompt with three simple constraints. That puts our measured cost at $3.08 per keeper of six seconds, which is $0.51 per delivered second, or $30.80 (£22.74) per finished minute in raw generation alone. It also puts our measured attempts per keeper at four, sitting between the practitioner figure of three and the congested figure of seven to eight.

Latency is the other cost nobody bills you for. Veo answered in about a minute, Seedance took eight minutes for the same six seconds. On a day of 60 attempts that is one hour of waiting against eight.

The rest of the invoice

Generation is the line item people quote because it is the one with a public price. It is rarely the largest.

A working commercial stack for one operator, at the tiers that actually grant commercial rights, runs like this: Kling Pro at $25.99, Runway Pro at $28, Midjourney Standard at $30, ElevenLabs Creator at $22 for 121,000 credits, and Premiere Pro at $22.99 on an annual plan. That is $128.98 a month, £95.24, before a single second is generated. Swap Premiere for DaVinci Resolve and it is free at the base tier, or $295 once for Studio. Topaz Video, which is what most people reach for when a 480p take is otherwise perfect, is $25 to $79 a month. Higgsfield's unlimited bundles are sold by the day, $99 for one and $250 for three, which tells you how the vendors expect crunch weeks to work.

Two traps sit in that list. Free tiers on Kling and elsewhere are non commercial, so a film made on one is not clearable. Suno and Udio grant commercial rights only on paid tiers, and after the WMG settlement what you receive is a license, not ownership of the master.

Published production numbers land where our per minute figure suggests they should. Across five documented AI shorts, totals ran $750 to $5,000, finished cost per minute ran $315 to $750, and schedules ran two to five days with teams of one to four people. One 70 second short was finished for $750, which works out at about $643 a minute. A three minute animated episode was finished for $950 by two people, about $315 a minute, and it is the cheapest per minute in the set because episodic work reuses characters and locations. At the other end, a five day sprint short with a team of four spent $5,000, and most of the difference is people rather than compute.

Compare that with the conventional indie tiers for 2026, where a $5,000 short film budget is itemized as $1,200 crew, $400 equipment, $800 post, $500 festival fees and $300 contingency. An AI short and a live action short at this level cost the same money. The difference is where it goes: almost none of it to people, and roughly a fifth to a third of it to compute that produced nothing usable.

What the community is actually arguing about

The cost conversation on X is thinner than the capability conversation, and worth reading for method rather than numbers.

@kimmonismus, in a post that reached about 23,000 impressions, and @rewind02 both push the same discipline: compare models on controlled parameters, same prompt, same duration, same settings, or the comparison is decoration. That is the methodology we used above, and it is the reason we can report a keeper rate at all. @rewind02 and @eCom_Amin, independently, describe Seedance 2.5 as the one they hand ad work to, which sits awkwardly beside our own result where Seedance missed the brief twice and took eight times as long. Our read: their workflows are image to video with a locked reference frame, ours was text to video, and prompt adherence is exactly where that difference shows up. @maverickecom, at roughly 93,000 impressions, and @MikeLeez69928 both report scaling brand UGC on generated video, which is the use case where a 25 percent selection rate stops mattering because volume is the product.

Worth saying plainly: none of these are edges on their own, and we only report the three claims that two independent accounts made separately. Nobody in that set published a cost per finished minute. That is the gap this piece exists to fill, and we would rather be corrected with someone else's measured run than agree with an estimate.

The number to plan against

If you are budgeting a generated short this quarter, plan on four to eight attempts per keeper, a quarter of your keepers surviving the edit, $30 to $60 per finished minute in generation at 1080p, and about $130 a month in subscriptions that you pay whether you shoot or not. Then add the cost of the model disappearing underneath you. The Sora 2 Videos API shuts down on 24 September, developers were notified on 24 March, and OpenAI's own migration table lists no recommended replacement for any of the six endpoints. Koosha moved a feature to Kling mid production for that reason.

There is one more cost we can measure but not price, which is the tax on your own judgment. Grading 60 near identical takes a day is the job now, and it is the reason two person teams beat four person teams on cost per minute in the published set. The cheapest line in the whole budget is still the one that reads "did not generate anything today". The most expensive is a pipeline built on an endpoint with a shutdown date.