Multimodal & Creative
Per Second, Per Image, Per Token: How Generative Video and Image APIs Actually Bill You
Video is billed per second and resolution, images per unit or pixel tier, and some models per output token. A guide to what those meters actually mean, and how to budget per usable asset instead of per second.
October 28, 2026 · 10 min read
Text models have converged on one meter: tokens in, tokens out. Generative media has not. Depending on the model, you are billed by the second, by the image, by the pixel tier, or by tokens derived from the output — and the number on the pricing page rarely maps onto the number on your invoice.
This is the segment where cross-vendor comparison goes wrong most often. Below is a map of the four metering models, a set of worked examples, and a budget method that survives contact with a production workflow.
Four metering models you will meet
1. Per second, multiplied by resolution. The video default. A base rate per second at a given resolution, with surcharges for higher resolutions and for an audio track.
2. Per second, in tiers. Two or three resolution buckets, each with its own per-second rate; no audio dimension.
3. Per image, sometimes per pixel tier. Flat price per generated image, or two price buckets split at a megapixel threshold.
4. Per output token. The newest and least intuitive: video or image generation metered as tokens, with different rates for different resolutions and whether an input video is supplied.
What the numbers look like
Illustrative only, rounded, and representative of published rates at the time of writing. Check live pricing before budgeting.
Model / family | Meter | Approximate rate |
|---|---|---|
Kling 3.x video | per second | ~ |
MiniMax H3 video | per second, tiered | ~ |
Grok Imagine Video | per second, tiered | ~ |
Seedance 2.x video | per output video token | ~ |
Seedream 5.0 Lite | per image | ~$0.0315 per image |
Seedream 5.0 Pro | per pixel tier | ~ |
Grok Imagine Image | per image | ~$0.014 per image |
Gemini image family | per output image token | ~$24 per million output image tokens, plus search tool charges |
A ten-second clip at 1080p with audio, at the rates above, lands around 3.43 — a 2.5x jump that has nothing to do with duration.
Why comparing headline rates misleads
- The units are not commensurable. A per-second rate and a per-token rate cannot be compared without knowing how many tokens a second of output costs on that specific model.
- Resolution and audio are multipliers, not options. A "cheap" model at its top resolution often costs more than a premium model at a mid resolution.
- Rejection rates differ. A cheaper model that needs four attempts to produce one usable asset is not cheaper.
- Input side charges exist. Reference images after the first, input video seconds, edit versus generate modes, and pixel-tier thresholds all add lines to the bill.
- Fixed overhead per request. A 5-second and a 20-second clip share the same submission overhead; short clips have worse effective cost per second.
- Attached tools are metered separately. Web search, image search, and other hosted tools appear as their own line items.
The unit that matters: cost per usable asset
Worked comparison for a 10-second marketing clip:
Path | Unit cost | Attempts | Cost per usable asset |
|---|---|---|---|
1080p with audio, direct | $1.37 | 2.5 | ~$3.43 |
720p silent exploration, then 1080p final | | 1 final | ~$4.13 |
4K, direct | $3.43 | 2.0 | ~$6.86 |
The middle row is the interesting one: it looks more expensive than the first row per attempt, yet it buys you four cheap explorations and a targeted final render. In practice the two-stage workflow — explore low, finish high — usually beats brute-forcing the expensive tier, unless your prompt library is already highly tuned.
Hidden input-side costs
Reference images: the first is often included; each additional image is billed per unit.
Input video: video-to-video is metered by input seconds, and on some models a supplied input video reduces the output rate — which changes how you should compare it.
Edit vs generate: editing an existing image is frequently a different line item than creating one.
Pixel tiers: the jump at thresholds like 2.61MP can double the per-image price. If your use case tolerates a smaller render, downscaling is a real discount.
Audio: either a separate line or a built-in surcharge on the higher resolution tier.
Budgeting practices that hold up
- Convert every rate to "per finished asset" before comparing anything. Per-second and per-token are intermediate units.
- Track attempts per accepted asset. This one number explains most budget surprises, and it is also your best quality signal.
- Set a hard budget per asset. When attempts exceed it, stop and route the prompt to a human instead of continuing to sample.
- Freeze seeds and templates. Random variation is how exploration budgets disappear; reuse a known-good composition seed when iterating on a detail.
- Use offline queues for batch work. Generation is slow; queueing it asynchronously is both cheaper and kinder to your users.
- Budget for failures. A timed-out render is still billed in most cases.
- Review pricing quarterly. This segment re-prices more often than text models, and resolution tiers get added rather than replaced.
A scoring card, not a vibe
To decide between models, run 20–30 prompts per family through a blind review with a fixed rubric:
Dimension | What you are actually judging |
|---|---|
Subject consistency | does the same character or product survive across frames |
Composition control | does the model respect framing and camera instructions |
Motion naturalness | physics, hands, text, reflections |
Text rendering | legibility if any on-screen text is required |
Usability | could this ship with minor retouching, yes or no |
Then divide total spend by accepted assets. That number — not the per-second rate — is what your finance team should see.
The takeaway
Generative media pricing rewards teams that think in finished outputs and punish those that think in unit rates. Choose the cheapest tier that can carry your exploration, spend your budget on the final render, and measure acceptance rate as carefully as you measure cost.
All prices are illustrative, rounded, and subject to change. Confirm current rates and billing rules with the relevant model pages.
