Cheapest Video Generation API: Cost per Second Compared

Every video API pricing page leads with the same number: dollars per second. It looks like the fastest way to find the cheapest option, and it is also the number most likely to mislead you. A rate of $0.02 per second sounds like a bargain until you throw away three clips out of four, or until the same model quotes $0.04 the moment you ask for 1080p.

This comparison lines up the published per-second rates, as of October 2026, for the models people actually call through an API: Veo 3.1, Kling v3, Wan 2.7, Seedance 2.0, Grok Imagine Video, P Video, LTX 2.3 Fast and a few others. Then it goes one step further and turns those rates into the number you can really budget with: cost per usable second.

💡 Short answer: the lowest listed rates sit between $0.005 and $0.05 per second, and they belong to P Video (draft mode and full quality), Grok Imagine 1.5 Lite, Grok Imagine Video and Veo 3.1 Lite. Premium tiers start at $0.40 per second, so Veo 3.1 at full price costs 20 times what P Video charges at 720p.

Prices move often. Treat every figure below as a snapshot, and confirm it on the provider's own page before you commit a budget.

Why Cost per Second Misleads

Per Second Is Not Per Clip

Not every provider bills the same way. Some charge per second of output video, some per million tokens, some in credits, and many resellers convert everything into a per-second figure with their own margin on top. One pricing tracker notes that ByteDance bills Seedance per million tokens rather than per second, which is why Seedance rows in most comparisons come from resellers.

Credits add another layer. Kling's own price list quotes Kling 3.0 in credits per second: 6 for silent 720p, 9 with audio, 8 for silent 1080p and 12 with audio. The dollar figures in this article are conversions of those credits, so the rate you pay depends on what a credit costs on your plan.

The same model can also carry different rates on different hosts. Wan 2.7 is listed at $0.086 per second at 720p on one reseller and $0.10 on two others, and at 1080p the spread runs from $0.144 to $0.15. Small gaps, but they compound across thousands of clips.

Audio and Resolution Move the Rate

Two settings change the rate more than the model name does. Kling v3 is listed at about $0.084 per second for 720p without native audio and $0.112 with it. At 1080p with audio the listing reaches $0.14, which is 67% more than the cheapest Kling configuration. Veo 3.1 Fast goes from $0.10 per second at 720p to $0.30 at 4K, a threefold jump for the same prompt.

Google's Veo prices include audio by default, so a Veo rate and a silent-clip rate from another provider are not the same product.

💡 Rule of thumb: compare rates at the same resolution and the same audio setting, or the table is comparing different things.

The October 2026 Price Table

Rates below come from first-party pages where they could be checked (Google's Gemini API pricing, xAI's docs, Pruna's docs), Kling's developer pricing converted to dollars, and reseller listings where the maker does not publish a per-second price. Reseller rows are marked with an asterisk.

Under $0.10 per Second

P Video is the price floor. At $0.02 per second for 720p, a five-second clip costs $0.10, and draft mode drops that to $0.025. Pruna positions draft as a quick preview pass: a 5-second 720p clip renders in about 2.5 seconds, against about 10 seconds at full quality.

Grok Imagine 1.5 Lite at a flat $0.02, Grok Imagine Video at $0.05 and Veo 3.1 Lite at $0.05 follow. The Veo Lite number deserves a second look because it includes audio, which several cheaper options price separately or skip. Kling v3 starts at $0.084 for silent 720p.

Between $0.10 and $0.15

This is the crowded middle. Veo 3.1 Fast sits at $0.10 to $0.12, Wan 2.7 lands around $0.09 to $0.15 depending on host and resolution, Hailuo 2.3 is listed at $0.10, and Kling v3 lands here once audio and 1080p are switched on. Seedance 2.0 Fast sits just under at $0.09 on a reseller.

For many teams this band is where quality starts to justify the price: steadier motion, closer prompt following, fewer clips in the bin. The cost-per-usable-second math further down puts numbers on that.

Premium Rates Above That

Veo 3.1 at $0.40 per second ($0.60 at 4K) is the premium benchmark on Google's own page. Inside Google's own lineup the ladder is steep: Veo 3.1 Lite at $0.05, Veo 3.1 Fast at $0.10 and Veo 3.1 at $0.40 for 720p, an 8x climb from bottom to top.

Seedance 2.0 at a reseller's $0.247 for its Pro tier and LTX 2.3 Pro at up to $0.32 for 4K fill out the top. Sora 2 is listed at $0.10 to $0.70 per second, though one tracker flags a September 24, 2026 sunset for OpenAI's API, so confirm availability before you build on it.

Premium rates make sense for the final hero shot, not for the thirty test renders that lead up to it.

What 500 Seconds of Video Costs

Per-second rates are easier to feel at volume. Here is the bill for 100 clips of 5 seconds each, 500 seconds in total, at 720p, assuming every single clip is a keeper:

The spread from $2.50 to $200 is an 80x gap on paper. It also assumes every clip is usable, which never happens in practice.

Now apply a realistic keep rate. If you publish only one clip in four, you need 400 generated clips, or 2,000 seconds, to end up with 100 usable ones. That multiplies every row by four: $40 for P Video at full quality, $100 for Veo 3.1 Lite and $800 for Veo 3.1. The ranking holds, but the absolute numbers become a different conversation.

Hidden Costs That Break Budgets

Failed and Unusable Generations

Most providers bill for a successful render and only a successful render. Google's pricing page, for one, says you are charged only if the video is successfully generated. But a render that succeeds technically and still shows a warped hand or a drifting logo is billed like any other. That is where the real waste lives, and the first prompt rarely lands.

Re-Renders and Upscales

Drafting at low resolution and re-rendering the winner at higher resolution means paying twice for the final shot. That is a good trade when the draft predicts the final, and a costly one when the composition shifts between passes. Lock a seed wherever the model allows it so the re-render stays close to the draft.

Queues, Limits, and Concurrency

Cost is not only money. PicassoIA's developer API allows 5 concurrent predictions per account, shared across all credentials and MCP connections, with prompts capped at 4,000 characters. Jobs are asynchronous: you create a prediction, poll it, then fetch the result. A cheap model behind a tight concurrency cap or a slow queue still burns developer hours, so check those limits on any provider before you commit.

Throughput is worth a quick calculation. At 5 concurrent predictions and about 110 seconds per render, the figure listed for a 720p Seedance 2.5 Lite example, the ceiling is roughly 160 clips an hour (5 x 3,600 / 110). Real times shift with load, but the ceiling tells you that a 500-clip batch takes about three hours at best.

Cost per Usable Second

The Formula

Cost per usable second = rate per second ÷ keep rate, where keep rate is the share of generated seconds you actually publish. If you trim clips in the edit, count only the seconds that survive the cut.

Two Worked Examples

These numbers are hypothetical. They show how the math behaves, not measured results.

• Cheap model, low keep rate: $0.02 per second with 1 clip in 4 kept works out to $0.08 per usable second.
• Pricier model, high keep rate: $0.05 per second with 3 clips in 4 kept works out to about $0.067 per usable second.

The model with the lower sticker price ends up costing 20% more. A draft workflow can flip the result again. Render eight 5-second drafts at $0.005 per second ($0.20 in total), then one 5-second final at $0.02 per second ($0.10). That is $0.30 for one usable 5-second clip, or $0.06 per usable second, provided the final matches the draft. Four straight full-quality attempts at $0.10 each would cost $0.40, or $0.08 per usable second.

Read across a row to see what waste costs you, and down a column to see how the model's price scales it. Two models break even when their rate-to-keep-rate ratios match, so a model priced 2.5 times higher has to keep 2.5 times as many clips just to tie.

How to Use Seedance 2.5 Lite

When the goal is iteration volume, the cheapest rate per second is zero. Seedance 2.5 Lite is described on its PicassoIA model page as free and unlimited for Wonder members, with clips of up to 10 seconds and synchronized audio. That rewrites the cost-per-usable-second math, because retries stop costing anything per clip.

Pick Your Settings

• Resolution: 480p for fast drafts, 720p for the sharper final. The model is limited to these two.
• Duration: 5 or 10 seconds.
• Aspect ratio: 16:9, 9:16, 1:1, 4:3, 3:4, 3:2 or 2:3. Leave the default match_input_image when you start from a photo.
• Audio: on by default through save_audio. Turn it off for a silent clip.
• Seed: lock it to reproduce a clip.

Write and Render the Prompt

1. Open the Seedance 2.5 Lite page and write a chronological prompt: subject, motion, camera move, lighting.
2. Optionally add an input image as the first frame. The clip inherits its aspect ratio. A last-frame image is also accepted, but only together with a first frame.
3. Render at 480p and 5 seconds until the composition works. The example renders listed on the model page took roughly 100 to 110 seconds for 5 and 10 second clips.
4. Lock the seed, switch to 720p for the final, and check the result, since a different resolution can change fine details.

Prefer a fixed format? Picasso IA Video renders 5-second clips at 24 fps, in 480p or 720p, with synchronized audio.

Calling It Through the API

The developer API is Replicate-style. The base URL is https://api.picassoia.com/v1, requests carry a Bearer token with the pia_sk_ prefix, you create a prediction with POST /v1/models/{owner}/{name}/predictions, and you poll GET /v1/predictions/{id} until it finishes. The models exposed there include picassoia/seedance-2.5-lite and picassoia/picassoia-video.

The docs describe API predictions as free for now, but the plan requirements are worded differently on the API page and the pricing page. Check both before you build a budget around "free".

Which API Fits Your Project

Social and Marketing Clips

Volume and speed matter most here. Draft on the cheap tier, with P Video at $0.02 or Grok Imagine 1.5 Lite, or iterate without a meter on Seedance 2.5 Lite and Picasso IA Video. When a clip needs audio and a cleaner finish, Veo 3.1 Lite at $0.05 per second is the budget option that includes it.

Product and Ad Production

Fewer clips, higher stakes. Draft cheap, then spend on Veo 3.1 Fast, Kling v3 or Wan 2.7 for the finals, and keep Veo 3.1 at $0.40 for the one hero shot that justifies it. Budget by cost per usable second, never by the sticker rate.

Run Your Own Cost Test

Published rates are a starting line, not a verdict. Your prompts, your subjects and your quality bar decide the keep rate, and the keep rate decides the real price.

Pick one prompt and run it on three models: one budget option such as Seedance 2.5 Lite or P Video, one middle option such as Veo 3.1 Fast, and one premium option such as Veo 3.1. Generate ten clips on each, count the keepers, and divide the spend by the usable seconds.

To keep the test honest:

1. Write the pass criteria before you generate: clean hands, readable motion, no flicker, audio that matches the action.
2. Use the same prompt, the same aspect ratio and the same resolution on every model.
3. Log the rate, the seconds generated and the seconds you would actually publish.
4. Divide total spend by usable seconds, and add any retry or re-render cost.
5. Repeat on a second prompt, because one prompt can flatter or punish a model.

Open Picasso IA, try the same prompt across the text-to-video models, and see which one gives you the lowest cost per usable second for your own content. One afternoon of testing will tell you more than any pricing table can.