AI Image Generation Cost vs Video: 2-9x More Per Frame

A second of 720p video from Seedance 2.0 costs $0.154 and contains 24 frames, which puts each frame at $0.0064^[1]. The cheapest single image on the same platform, Nano Banana 2 Lite, costs $0.015^[2]. The AI image generation cost is 2.3 times the video frame, and the video frame arrives with 23 siblings that are temporally consistent with it.

That inversion holds across every tier we sell and gets wider at the cheap end, where a Seedance 2.0 Mini frame runs $0.0015 against $0.015 for the least expensive standalone image. Ten to one. It is worth understanding why, because it changes which tool you reach for when you need a still.

TL;DR

• A 720p video frame costs $0.0064; the cheapest standalone image costs $0.015^[1][2].
• Normalized per megapixel, images run 2 to 9 times the AI image generation cost of a video frame.
• The AI image generation cost gap is widest at the low end (Mini 480p frame at $0.0015, roughly 10x cheaper than any image) and narrowest at 4K, where images are about 2x.
• Video pricing meters on total pixels across the clip; image pricing meters per generation, so short clips subsidize their own frame count.
• The catch is control. A video frame is not addressable, not re-rollable, and not prompt-steerable in isolation.
• Extracting stills from generated video is cheap and legitimate; it is also the wrong tool when you need nine unrelated compositions.

The per-frame numbers

Video is quoted per second and renders at 24 frames per second, so per-frame cost is the per-second rate divided by 24. Across the Seedance 2.0 family^[1][3]:

Against the image catalog^[2][4][5][6]:

The AI image generation cost per megapixel bottoms out at $0.0143 and the cheapest video tier costs $0.0035. Four times. Compare the flagship image models against mid-tier video and the multiple reaches nine.

Why AI image generation cost diverges from video

Video meters on total pixels shipped. The published formula for the Seedance family multiplies duration by frame area by frame rate, then divides into token buckets^[7]. Every frame is priced identically to every other frame, and there is no fixed per-request overhead worth speaking of.

Image generation prices per invocation. Whatever the model spends on prompt interpretation, sampling schedule and safety passes gets amortized across exactly one output. A four-second clip amortizes the same class of overhead across 96.

The second factor is what the model is being asked to do. An image model treats every generation as an independent problem: fresh composition, fresh subject, fresh everything. A video model solves one composition and then propagates it, which is a fundamentally cheaper operation per frame even though it is a harder problem overall.

Neither of those is a pricing quirk anyone chose. It is what the workloads actually cost to run, surfaced in the rate card.

When pulling a still from video is the right call

The arithmetic suggests an obvious arbitrage, and for some jobs it genuinely works.

If you need a hero shot of a specific subject in a specific pose, generating two seconds of 1080p video costs $0.766 and returns 48 candidate frames at $0.016 each. Pick the best. Against $0.033 for a single Nano Banana Pro image, you are paying 23 times the unit price for 48 times the options, with the subject held consistent across all of them.

If you need a character to appear in several shots with the same face, video is the cheaper consistency mechanism by a wide margin. That is the whole premise of reference-driven generation, and it is why the reference tier exists at a lower per-second rate.

If you need contact-sheet variety on one concept, the same logic applies. Twenty seconds of Mini 480p is $0.72 and yields 480 frames.

When it is the wrong call, which is often

A video frame is not addressable. You cannot prompt frame 37. You steer the clip and accept what lands, whereas an image model gives you a seed, a full prompt, and a re-roll on a single output.

Compression is real. Video frames arrive from an encoded stream with motion-compensated artifacts that a still generator never introduces. At 480p, on a subject with fine texture, this is visible and not fixable in post.

Text rendering falls apart. If the still needs legible words in it, GPT Image 2 exists precisely because video models do not hold typography stable frame to frame^[6].

Aspect ratios are constrained. Video tiers ship 16:9 and a handful of standard ratios. Image models on the same platform reach 1:4, 4:1 and 8:1^[4], which video cannot do at any price.

And nine unrelated compositions means nine generations either way. The per-frame advantage evaporates the moment you need frames that are not related to each other, because relatedness is the entire thing you are buying.

The number that actually matters for a product

Neither per-frame nor per-image AI image generation cost is the figure to budget on. Budget on cost per accepted output.

An image workflow accepting one in four generations at $0.030 costs $0.120 per keeper. A video workflow that generates two seconds at 720p for $0.308, then keeps one frame, costs $0.308 per keeper, and that is worse, right up until you need the second keeper from the same clip, at which point it halves.

Run your own acceptance rate before deciding which AI image generation cost applies to you. The AI image generation cost advantage of video frames is real, but it only converts into money saved when your workflow consumes more than one frame per generation.

Where the gap shows up on a real bill

Take a storefront that needs 500 product stills a month. At Nano Banana Pro, the AI image generation cost is $16.50. Harvesting the same 500 frames from 1080p video means 21 seconds of generation at $8.04, roughly half, and every frame shares lighting and camera treatment.

Now change the brief to 500 unrelated products. The video route collapses, because 500 different subjects means 500 separate clips, and the AI image generation cost is suddenly the cheaper option by a wide margin. Same monthly volume, opposite answer, and the deciding variable was never price.

That is the useful way to read the tables above. The AI image generation cost premium is not a markup on equivalent output. It is what you pay for independence between generations, and independence is either the entire point of your workflow or it is dead weight. Work out which one you have before you optimize the wrong number.

FAQ

Does a video frame match the quality an AI image generation cost buys?

No. It comes out of an encoded stream, so it carries compression artifacts, and it is generally softer on fine texture. At 1080p and above the difference narrows considerably.

Can I extract frames from a generated video?

Yes. The output is a standard video file, so any frame extraction tool handles it. Nothing on the platform restricts it.

Why is 4K video cheaper per megapixel than 1080p?

The token rate for the 4K tier is lower than the 1080p tier^[7]. The total cost is still far higher because the frame carries four times the pixels, but normalized per megapixel it comes out ahead.

Which is cheaper for character consistency across shots?

Video, decisively. Reference-driven generation keeps a subject stable across frames at the per-second rate, which is what image models charge per output to approximate.

Does the reference video tier change this math?

It lowers the per-second rate but adds your input clip to the billed duration^[7]. For frame-harvesting with no reference input, the plain rate is what applies.

What resolution do the video tiers actually render?

480p renders around 864 x 496, 720p at 1280 x 720, 1080p at 1920 x 1080 and 4K at 3840 x 2160^[7]. Those are the dimensions the per-megapixel column uses.

Do failed generations cost money?

No. Failed jobs refund automatically on reAPI^[1].

Choosing between the two

The rate cards say video frames are cheap and AI image generation cost is high, and taken literally that is correct by a factor of two to nine. Taken as advice it is misleading, because the two products are not substitutes. Video sells you a sequence in which every frame relates to its neighbours. Images sell you independent control over one output at a time.

The workflows where the math genuinely pays are the ones that want relatedness: character sheets, shot variations on a fixed subject, contact sheets from a single concept. For anything requiring nine distinct compositions or legible text, the higher AI image generation cost buys control you cannot get from a video frame at any price, and that control is usually the cheaper purchase in the end.

References

1. reAPI. Seedance 2.0 — model page and live pricing. Retrieved August 2026 from reapi.ai/models/seedance-2-0
2. reAPI. Nano Banana 2 Lite — model page and live pricing. Retrieved August 2026 from reapi.ai/models/nano-banana-2-lite
3. reAPI. Seedance 2.0 Mini — model page and live pricing. Retrieved August 2026 from reapi.ai/models/seedance-2-0-mini
4. reAPI. Nano Banana 2 — model page and live pricing. Retrieved August 2026 from reapi.ai/models/gemini-3-1-flash-image-preview
5. reAPI. Nano Banana Pro — model page and live pricing. Retrieved August 2026 from reapi.ai/models/gemini-3-pro-image-preview
6. reAPI. GPT Image 2 — model page and live pricing. Retrieved August 2026 from reapi.ai/models/gpt-image-2
7. BytePlus. ModelArk — model pricing, token calculation formula and per-video examples. Retrieved August 2026 from docs.byteplus.com/en/docs/ModelArk/1544106