Every creator who has priced AI video knows the pattern by now. You find a model you like, you read the rate card, and the number is quoted per second. A ten-second clip is rarely a one-off purchase — it is a draft you will throw away, a variation you want to see, a second pass at the lighting. Cost per second decides how many of those attempts you can actually afford.
That is why pricing matters more in this category than it did a year ago. The quality gap between the leading models has narrowed enough that most teams can get a usable shot from several of them. What still separates the platforms is arithmetic.
The market is expanding in step with that demand. Grand View Research estimates the global AI video generator market at $788.5 million in 2025 and projects it to reach $3.44 billion by 2033, a compound annual growth rate of 20.3%. More tooling, more models, and more creators comparing the cost of a shot before they commit to it.
What the official rates actually look like
Start with the reference prices, because they set the floor for everything else.
MiniMax publishes $0.08 per second for H3 at 768P and $0.13 per second at 2K on its official pricing page. Seedance 2.0 through fal.ai lists at $0.3034 per second at 720P, $0.682 at 1080P, and $0.2419 for the Fast tier at 720P. OpenRouter quotes Seedance 2.0 Mini at $0.03363 per second at 480P and $0.0756 at 720P.
None of those numbers are unreasonable for what the hardware is doing. They are simply numbers that multiply. A single ten-second draft at fal.ai's 1080P Seedance rate is $6.82. Run five variations before you pick one and you have spent $34 on a clip that may still need a reshoot at a different resolution.
Reference route
Rate
10-second clip
Seedance 2.0 · 1080P (fal.ai)
$0.682/s
$6.82
Seedance 2.0 · 720P (fal.ai)
$0.3034/s
$3.03
Seedance 2.0 Fast · 720P (fal.ai)
$0.2419/s
$2.42
MiniMax H3 · 2K (official)
$0.13/s
$1.30
MiniMax H3 · 768P (official)
$0.08/s
$0.80
Seedance 2.0 Mini · 720P (OpenRouter)
$0.0756/s
$0.76
Seedance 2.0 Mini · 480P (OpenRouter)
$0.03363/s
$0.34
The difference between the top and bottom of that table is a factor of twenty, for clips of similar length. That spread is what people are actually shopping for.
Where the same models cost less
Callirra takes the same model families and quotes them against a single credit balance, with the full cost shown before a generation starts. Public credits are straightforward: one credit equals one cent.
MiniMax H3 is the clearest example. Callirra runs it at $0.01 per second in 480P, $0.02 at 768P, $0.03 at 1080P and $0.10 at 2K:
H3 output tier
Callirra rate
Official MiniMax rate
Difference
480P
$0.01/s
—
—
768P
$0.02/s
$0.08/s
75% lower
1080P
$0.03/s
—
—
2K
$0.10/s
$0.13/s
23% lower
A ten-second H3 clip at 768P costs twenty cents. That is the tier most people will live in while they iterate, and it changes the economics of experimenting: ten variations cost two dollars instead of eight.
Seedance 2.x: two routes instead of one
Seedance deserves a separate look, because the way it is offered matters. Callirra labels two routes rather than hiding the difference behind a single card.
The everyday route is tuned for frequent generation and tight instruction following, with real-face reference support and a low per-second quote:
Model
480P
720P
1080P
4K
Seedance 2.0 Mini
$0.02/s
$0.04/s
—
—
Seedance 2.0 Fast
$0.06/s
$0.13/s
—
—
Seedance 2.0
$0.10/s
$0.21/s
$0.52/s
$1.05/s
Seedance 2.5
$0.14/s
$0.32/s
$0.56/s
—
Matched against the public references above, Seedance 2.0 Mini comes in roughly 40% below the OpenRouter rate at 480P and about 47% below it at 720P, while Seedance 2.0 Fast at 720P is around 46% lower than the fal.ai figure.
The second route is the extreme-width lane. It exists for prompts and reference inputs that need a wider review setting, and it raises the prompt ceiling to 20,000 characters. It is a genuine capability choice rather than a fallback, and the price stays visible before you generate:
Model
480P
720P
1080P
4K
Seedance 2.0 Mini wide
$0.03/s
$0.06/s
—
—
Seedance 2.0 Fast wide
$0.09/s
$0.18/s
—
—
Seedance 2.0 wide
$0.14/s
$0.30/s
$0.74/s
$1.51/s
Seedance 2.5 wide
$0.21/s
$0.46/s
$1.15/s
—
The practical rule is simple: stay on the ordinary route while you are refining a shot, and switch to the wide route when the prompt itself is the constraint rather than the budget.
The image side works the same way
Image generation has its own price ladder, and the same per-output logic applies. GPT Image 2.5 runs across three resolutions and five quality tiers, so you can trade fidelity against cost shot by shot:
Resolution
low
medium
high
xhigh
max
1K
$0.03
$0.04
$0.06
$0.09
$0.13
2K
$0.05
$0.07
$0.10
$0.15
$0.22
4K
$0.08
$0.11
$0.16
$0.24
$0.35
Nano Banana is where the gap against the reference rate is widest. Comparing the same output tiers against the equivalents calculated from Google's published Gemini API rates:
Model
Tier
Callirra
Google reference equivalent
Difference
Nano Banana 2
1K
$0.04
$0.067
about 40% lower
Nano Banana 2
2K
$0.05
$0.101
about 50% lower
Nano Banana 2
4K
$0.06
$0.151
about 60% lower
Nano Banana Pro
1K
$0.06
$0.134
about 55% lower
Nano Banana Pro
2K
$0.07
$0.134
about 48% lower
Nano Banana Pro
4K
$0.08
$0.24
about 67% lower
Seedream fills the high-resolution end of the shelf: Seedream 5 Lite runs at $0.04, $0.05 and $0.06 for 2K, 3K and 4K, and Seedream 5 Pro at $0.06 and $0.12 for 1K and 2K.
Why per-second pricing changes how you work
Once cost is visible before every run, the workflow changes in ways that are easy to underestimate.
Draft cheap, deliver expensive. A 480P H3 pass at one cent per second is a genuine storyboard tool. Decide on framing there, then render the approved version at 1080P or 2K.
Use references instead of retries. MiniMax H3 accepts up to nine reference images and up to three reference audio clips, with first and last-frame conditioning. Better inputs reduce the number of attempts, which is where the money actually goes.
Let a prompt assistant do the first pass. Context-IR expands a short idea into a structured subject, action, camera, lighting and audio brief that you preview before applying. The original wording stays available, and the preview is priced separately from the render.
Measure in credits, not in feelings. Because one credit equals one cent, a shot's cost translates directly into a number you can put in a spreadsheet next to its resolution and duration.
One balance across video, image and prompts
Model catalogs tend to sprawl across subscriptions. Callirra keeps video, image, audio and prompt tools on one balance, with plans from $9.90 per month for 800 credits up to $149 per month for 18,000, and yearly billing bringing the effective price to roughly $0.0066 per credit. One-time packs start at $5.90 for 500 credits, and weekly free credits cover the first experiments.
The same catalog and the same quotes are available from the web generator, the HTTP API, the CLI and the MCP server, so a prototype built in the browser does not need to be re-priced when it moves into a pipeline. Image Prompts Studio and the reusable prompt libraries shorten the distance between an idea and a first result.
Pricing in this category will keep moving as new models arrive and hardware gets cheaper. What will not change is the arithmetic that decides how many attempts you can afford. Pick the route, check the quote, and spend the budget on iterations rather than on the first attempt.
