Most companies that “add AI video” to their marketing in 2026 do it the way they added stock photography in 2010: someone on the team signs up for a tool, pays a monthly fee, and the cost disappears into a card statement. That works until the first quarter where video becomes a real channel and the finance team asks what a clip actually costs. The answer turns out to be surprisingly hard to give, and the reason is structural, not a matter of reading the pricing page more carefully.
One model, six prices
The dominant generative video models, Seedance 2.5 from ByteDance, Wan 3.0 from Alibaba, and Google’s Veo line, are licensed to dozens of downstream platforms. Each platform wraps the same model in its own currency (credits, tokens, seconds, “generations”), its own tiers, and its own rounding rules. The result is that a business buying AI video is rarely comparing models; it is comparing packaging. We normalized the offers of fifteen platforms that resell Seedance 2.5 to a single unit, the price of one ten-second 1080p clip with sound, and the spread was almost exactly ten to one: from about fifty cents on the model vendor’s own consumer app to five dollars on the most expensive aggregator. The full table, with sources and the arithmetic for each plan, is in our Seedance 2.5 pricing study across 15 platforms.
Three things drive the spread, and none of them is “some vendors overcharge”:
| Layer | Effective price per 10-s clip | What the margin pays for |
|---|---|---|
| Vendor’s own app | $0.50–1.00 | Customer acquisition for the model owner; no API, limited features |
| Developer API | ~$1.00–1.20 | The wholesale rate; you build the product around it |
| Workflow platforms | $1.15–2.00 | Characters, storyboards, editing, consent and rights tooling |
| Multi-model aggregators | $2.00–5.00 | One key for many models; convenience premium |
Read as a strategy question, the table says: the model itself is a commodity at roughly a dollar a clip, and everything above that line is a product decision about what workflow you want to buy versus build.
The unit that matters is not the clip
For a marketing team the relevant unit is a published asset: one ad variant, one product explainer, one social post. A published asset almost never equals one generation. Across the campaigns we have run and observed, a usable ten-second ad takes between three and eight generations once you count rejected takes, reframes and a second pass for sound. That multiplier is the number to put in the model, and it is where the platforms differ far more than in list price.
- Text-to-video only platforms sit at the high end of the multiplier, because every retry re-rolls composition and faces.
- Platforms with first-frame and reference workflows sit at the low end, because the expensive video call is only made after a cheap image is approved.
- Platforms with persistent characters collapse the multiplier further for series content, because the same cast is reused across a month of assets without re-prompting.
A tool that charges 30 % more per clip but halves the retries is cheaper. Very few procurement comparisons capture this, because the pricing page does not.
Build, buy, or bundle: a decision rule
We see three viable postures, and the right one depends on volume and on whether video is a product feature or a marketing input.
- Buy a subscription if you publish fewer than roughly two hundred assets a month and video is a marketing input. The workflow tooling is worth more than the margin it costs you, and the alternative is an engineer’s week you will not get back.
- Build on the API if video is inside your product, you have engineers, and volume is in the thousands of clips. At that scale the dollar-a-clip floor and the ability to tune retries yourself pay for the integration in a quarter.
- Bundle if you are an agency: buy one workflow platform for the team and pass through per-client token budgets. The reporting is cleaner than a dozen personal subscriptions and the persistent-character feature is what lets you sell “the same brand ambassador every month” as a retainer.
The hidden line items
Whichever posture you choose, four costs sit outside the per-clip price and routinely double the real number:
- Resolution and sound surcharges. 1080p with generated audio commonly costs two to four times the 720p silent price on the same platform. Most quoted prices are for the cheap tier.
- Rights and consent. If real people appear, someone has to obtain and store consent. Platforms that build a liveness check and a consent record into the workflow are pricing in a legal cost you would otherwise carry yourself.
- Storage and hosting. Generated video is large. Plans that include a gigabyte of storage are effectively bundling a hosting bill.
- Promotional pricing. Model vendors run launch discounts of 30–50 % that expire on a date buried in the docs. Any cost model built in a promo window should carry the post-promo price as the base case.
What to put in the board deck
A defensible AI video budget in 2026 has three numbers: the normalized cost of one clip at the resolution you actually publish, the observed generations-per-published-asset multiplier for your team, and the monthly asset count. Multiply them, add the hidden line items, and compare that figure against what the same assets cost with a crew and an editor. For most small and mid-sized marketing teams the comparison is not close, which is why the interesting question has stopped being “should we use AI video” and become “which layer of the stack should we pay for”. The pricing study linked above is a good place to start on the first number; the second one you can only measure by running a month.