AI Video API Cost in 2026: List Price vs Usable-Clip Cost

Nobody overspends on video generation because they picked the wrong rate card. They overspend because they budgeted for the clips they wanted and paid for the clips they threw away. The per-second rate is the smallest term in the bill. The term that moved in 2026 is how many generations it takes to get one usable shot, and it roughly tripled with no price change at all.

The formula:  effective cost = list rate x seconds x attempts per usable clip
What moved:   attempts per keeper went 2-3 -> 7-8 during 2026 congestion (36kr, 2026-04-01)
What did not: the rate card
Measured:     6s 480p Seedance 2.5 billed $0.66 in 103s; 30s billed $3.30 in 200s
Delivered:    that same 6s clip is $1.32 at 2 attempts and $5.28 at 8
Expiring:     ByteDance's 1080p launch discount window closes 2026-09-17; gateways track it
Snapshot:     2026-08-26

Everything below is either a rate card we read on 2026-08-26, a generation we ran and were billed for, or reporting we cite by name. Where the number people repeat does not survive a check, we say so rather than repeating it.

Why Is the Rate Card the Wrong Number?

Because it prices generations, and you ship keepers.

Every video API bills per output second of whatever comes back, including the takes you delete. So the unit that matters is not the per-second rate but the per-second rate multiplied by your reject ratio:

effective_cost_per_delivered_second = list_rate x (generations / accepted)

That second factor is the one nobody puts in a pricing comparison, because no vendor can quote it. It depends on the model, your prompt discipline, how specific your brief is, and, uncomfortably, on how busy the provider is that week.

Here is what that does to a real bill. We generated a 6-second 480p clip on Seedance 2.5 on 2026-08-13 and were billed $0.66:

The spread between the top and bottom of that table is 8x. The spread between the cheapest and most expensive rate card on the market for the same model is under 3x. If you are optimising the rate card while ignoring the attempt count, you are tuning the smaller dial.

What Actually Changed in 2026?

The attempt count, and it changed for reasons outside your pipeline.

This is the part with real reporting behind it. On 2026-04-01, 36kr documented what Seedance users were seeing: where creators had previously generated two or three clips and picked one usable take, they were now often generating seven or eight to find one. Same prompts, same model name, roughly triple the burn.

The surrounding conditions were not subtle, and they come from three separate reports that get blurred together in circulation. The 36kr piece quotes a creator who submitted a prompt and was told 80,000 people were ahead of them, and it attributes the improving wait times to deliberate quality throttling rather than added capacity, the practice Chinese coverage named 降智, roughly “dumbing down”. A 2026-04-28 write-up by 白鲸实验室, published on bianews, records one task sitting 68 hours in queue with the position counter climbing from 1,900 to past 50,000, and afternoon queues into six figures. 21jingji covers the consumer side in the same window: the early-adopter 40%-off tier withdrawn, and the credit cost of a 15-second generation for premium members going from 45 to 120, about 167%.

You do not have to accept any editorial framing to use this. The operational lesson stands on its own: your provider’s capacity state is an input to your unit economics, and it is not on the rate card. A quarter where attempts per keeper drifts from 3 to 6 doubles your content cost while your finance dashboard shows a flat price per second.

That is also why a cost model with a single hardcoded multiplier is worthless. Instrument the ratio and recompute it weekly.

What Do the Rate Cards Actually Say?

Read on 2026-08-26, for the same model, so the comparison is real.

Four things in that table are worth more than the cheapest cell.

The first is that the same model is not the same product on every route. ModelArk and Volcano Engine Ark are separate consoles with separate model IDs, dreamina-seedance-2-5-260628 internationally and doubao-seedance-2-5-260628 in China, and you cannot mix credentials or IDs between them.

The second is that credit pricing is rate pricing wearing a costume. Runway’s API credit is one cent, so 30 credits per second is thirty cents per second and 68 at 1080p is sixty-eight, before the surcharge it adds for each second of input or reference video you supply. That puts Runway’s 1080p rate 42% above the gateway’s discounted $0.48 and 13% above its $0.60 list. Higgsfield’s credits only convert through a subscription, which means its effective rate depends on how much of your monthly allowance you actually burn. A plan you use at 40% is a rate card 2.5x worse than the one advertised.

The third is that token billing and per-second billing answer different questions. ByteDance bills its own API per token, which is honest to how the model works and useless for forecasting a 25-shot episode until you have run the token formula against real prompts. Per-second gateway pricing is a markup on that, bought for predictability.

The fourth is the one with a date attached, and it is below.

What Did We Actually Get Billed?

Three runs on 2026-08-13, and the numbers are less flattering than the rate card implies.

Billed is what the gateway actually charged, discounts included. Seedance 2.5 at 480p carries no standing discount, so $0.66 is the $0.11/s list. Seedance 2.0 carries 10%, so $0.378 is $0.063/s against a $0.07/s list.

The last row is the good news and it is worth building for. When Seedance 2.5 was still capped at 720p, a 1080p request returned resolution 1080p not supported; allowed: [480p 720p] rather than quietly downscaling and billing you. A synchronous 400 costs nothing. Any parameter you can get rejected at submit time is a parameter that cannot become a wasted generation, so validate aggressively and prefer providers that fail loudly.

The 2.0 versus 2.5 pair is the sharper lesson. Seedance 2.0 lists 36% below 2.5 at 480p, 43% below once its standing discount lands, and took 2.8x longer to return. If 2.0 also needs one extra attempt on your shot list, its price advantage is gone and you have paid for the privilege in latency. We wrote up where each model still wins in Seedance 2.5 vs 2.0.

How Do You Actually Cut the Bill?

Attack attempts. The rate card has maybe 30% of movement in it; the attempt count has 400%.

Four levers, in the order we would apply them to an existing pipeline.

Draft cheap, deliver expensive. At $0.11 against $0.48, 480p costs less than a quarter of 1080p on the same model. Iterate the composition, the beats and the camera at 480p, then re-render only the approved shot at delivery resolution. A 25-shot episode that iterates four times at 480p and renders once at 1080p costs meaningfully less than one that iterates twice at 1080p, and it is the same finished episode.

Remove the model’s freedom to reinterpret. Every degree of freedom you leave open is a coin flip you pay for. Reference assets pin identity and setting; Seedance 2.5 accepts up to 50 of them as 30 images, 10 videos and 10 audio clips. First and last frames pin where a shot starts and ends, and in our first and last frame test the 2.5 last-frame error was 5.4 against 35.9 on 2.0 Mini. Integer-second timestamps pin when beats land. None of these lower the rate. All of them lower how often you re-roll.

Get your task type right the first time. On Seedance 2.5 the task is inferred from words in your prompt, so an edit request without an edit trigger word is classified as a reference-to-video job and returns a brand new clip. It is a perfectly good clip. It is also a full billed generation that answered a question you did not ask. Setting the task type explicitly moves that failure to submit time, where it is free. The trigger words and the failure modes are in how to use Seedance 2.5.

Measure the ratio, not the average. Log billed amount and generation count per accepted shot, not per request. A per-request average tells you the rate card, which you already knew. Cost per accepted shot tells you when your provider’s capacity state has quietly moved, usually two weeks before anyone notices in the finance review. If you are polling asynchronously, the same log gives you queue behaviour for free; we broke that pattern down in video generation API polling.

What Expires on 2026-09-17?

ByteDance’s 1080p launch discount, and every rate that tracks it.

Seedance 2.5 shipped without 1080p and gained it on 2026-08-17 Beijing time. ByteDance attached a launch discount running to 2026-09-17, and the gateways mirrored it at various depths: 20% off at ofox and Atlas Cloud, 28% off at EvoLink, 72% of list on Volcano Engine Ark. Different depths, one upstream expiry. Most of those gateways publish the discounted number without publishing the date, so the expiry will not be visible on the page you are pricing from.

If you are building a cost model this month, carry both the promotional and the post-promotional 1080p rate, because a spreadsheet anchored on $0.48 per second silently becomes wrong on September 17th and nothing in your pipeline will tell you. On a 25-shot, 60-second episode delivered at 1080p, the difference between $0.48 and $0.60 per second is $7.20 per episode before attempts, and $28.80 to $57.60 after a realistic multiplier. At 1,000 episodes that is a budget line, not a rounding error.

Which Numbers in Circulation Survive a Source Check?

Two of the three do not, and the third is real but routinely pinned on the wrong outlet.

Cost claims travel further than their sources. Here is what we found when we chased the three that circulate hardest on this topic.

A widely quoted V2EX post said to be titled along the lines of “it’s not the model, it’s your budget” does not appear to exist under that title. The closest real thread, a June discussion of pricing, failure and waiting, argues that unpredictable pricing hurts more than model quality does, which is a fair point, but it contains no cost figures at all. The commonly repeated per-clip and per-ad figures attached to it have no traceable origin.

A per-episode figure of roughly $7.6 for a 60-second, 25-shot production is often attributed to Tencent Cloud. Tencent Cloud’s own pipeline write-ups publish cost reduction percentages and per-task unit costs, such as storyboard splitting at CNY 2.8 against CNY 9, and production timings of about three hours per 2-3 minute episode. They do not publish a total per-episode cost. The $7.6 may be someone’s arithmetic; it is not a vendor figure.

The queue numbers are the opposite problem. They are real, and they get attributed to whichever outlet the writer happened to have open. The 80,000 that circulates most widely is a creator’s own account, quoted in that 36kr piece on 2026-04-01. The climb from 1,900 to past 50,000 and the 68-hour wait are neither 36kr’s nor 21jingji’s; they are 白鲸实验室’s, published on bianews on 2026-04-28, which is also where the six-figure afternoon queues come from. The 21jingji article usually cited for them describes a six-hour wait and forced quality throttling and contains neither figure. Three reports, three different weeks. Quote each with its outlet, or you are drawing a trend line through snapshots that were never on the same axis.

If you are building a business case on any of these, build it on your own logs instead. The one number that matters is one only your pipeline can produce.

The Short Version

The rate card is a starting point that vendors compete on and a term you have almost no leverage over. Attempts per usable clip is the term that moved 3x in 2026, that varies with your provider’s capacity state, that responds to reference assets and explicit task types and resolution laddering, and that nobody will quote you.

Price it, log it, and recompute it weekly. Then compare rate cards, with the multiplier already in the sheet.

Two maps sit behind this arithmetic: the Seedance API guide for per-second rates by tier, and six costs the rate card does not show for the text-model equivalent.