How much does AI video cost per second through an API?

Through an API, AI video generation is priced per second of output on most models, so the cost is the rate times the clip's length, and the model and resolution set the rate. On Sume the billed rate (the provider's list price × 1.25) runs from $0.0125 a second on grok-imagine-video-1.5 to about $1.43 a second on seedance-2.5 at 1080p, plus a 5.5% agent fee by default; a 10-second 720p clip from a prompt or a photo costs $0.13 to $5.82.

Other video APIs also price per second. Kling's API lists Kling 3.0 at $0.084 to $0.42 a second, and Runway's API lists its video models at 5 to 150 credits a second, at $0.01 per credit. A dated table of Runway, Kling, fal, and Replicate prices follows Sume's.

The Sume rates below are computed from the provider list prices in Sume's pricing code and the × 1.25 billing rule in the Video generation docs, read on 2026-09-27. For live rates, GET /v1/videos/models lists each model's pricing_skus, rounded to four decimals.

How much does AI video cost per second on Sume?

Sume bills every video model at the provider's list price × 1.25 and reserves the amount from your workspace balance when you submit. Most models charge per second of output, and most of those charge more per second at a higher resolution. The Seedance models charge per 1,000 video tokens instead.

What do Runway, Kling, fal and Replicate charge per second?

Each vendor's own pricing page, read on 2026-09-28. The USD column is only filled where the page states the dollar amount or the credit price: Runway sells credits at $0.01 each, and Kling prints the dollar price beside each unit count.

• Runway also adds charges on some models: seedance2_5 bills input and reference video per second and has an 80 credit minimum per generation, and grok_imagine_1_5 adds 1 credit per image or audio reference.
• Kling's table is its resource package deduction rules for video generation. Kling 3.0's per-second price rises with resolution, and at 720p and 1080p also with native audio.
• fal bills video models per second or per video, depending on the model. Replicate bills most models by the time they take to run and some by input and output, with an estimate on each model's page.
• Sume bills its provider's list price × 1.25, plus the agent fee; one key and one balance cover every model in its catalog.

Can one API key cover several of these models?

Yes. Sume's POST /v1/videos calls every model in its catalog with one key and one balance; fal's pricing page lists several video models with their own prices, and Replicate's says it hosts thousands of models. One API for Kling, Seedance and other video models covers what that bill looks like, and fal.ai and Replicate alternatives for video compares the platforms.

How are Seedance clips priced?

By video tokens, billed per 1,000. Sume's pricing code currently counts them as width × height × seconds × 24 ÷ 1,024, with a 720p 16:9 frame at 1280×720, so each second is 21,600 tokens, about $0.58 a second on seedance-2.5. Other aspect ratios have slightly different pixel counts, which is why the top of the 720p range above is a 4:3 or 3:4 clip.

Resolution raises both factors on seedance-2.5: a 1080p frame has more pixels, and 1080p has its own, higher per-token rate. A 10-second 1080p 16:9 clip reserves $14.22.

How much does a minute of AI video cost?

No single clip runs a minute. seedance-2.5 and wan-3.0 stop at 30 seconds, and every other model at 15 seconds or less, so a minute of footage is several clips; see AI video length limits by model. Sixty seconds of output costs $0.76 on grok-imagine-video-1.5 (four 15-second clips), $7.50 on wan-3.0 at 720p (two 30-second clips), and $22.50 on gemini-omni-flash-1.1 at 4K (six 10-second clips), before the fee.

Joining the clips into one MP4 with Timeline 1.0 is listed at $0.10 per output minute on API pricing; How to assemble a long-form video shows the render.

Does image-to-video cost the same as text-to-video?

Yes, for the same model, length, resolution, and aspect ratio. A model's rates in pricing_skus vary by resolution, or by sound on kling-3, and none of them depends on whether the clip starts from a prompt or from your photo as the first frame.

References are where extra cost can appear. For a reference video on the Seedance models, Sume's current estimate adds assumed input seconds to the token count; Why are AI videos so expensive? covers that and the other cost drivers.

What else changes what I pay?

• The agent fee: 5.5% of the model amount by default, explained in What is the Sume agent fee?
• Failed jobs: the reserved amount is refunded on failure, or on cancellation before capture. See Do failed AI video jobs cost money?
• sume/auto: Sume picks the model and does not disclose which. Read the job's usage.cost, the Sume billable amount, when it completes.
• Checking first: Estimate AI video cost before you run lists the ways to see a price before paying, and AI video cost per run reads what a finished job cost.