AI Video Generator API Pricing (2026): Per-Second Rates Side by Side, the Day Before Sora Leaves the API

If you are reading this on September 23, you have one day left on the Sora 2 API. OpenAI's own deprecations page lists the Videos API and five Sora 2 model names shutting down on September 24, 2026, and the column labelled "Recommended replacement" still shows three dashes on every row. So the question has changed from "is Sora worth it" to "what does the same clip cost somewhere else, today".

Most of the pages ranking for this question quote a price that is weeks old, or someone else's resale price. This one does something narrower and more useful: every number below was read from the vendor's own pricing or developer documentation page on 2026-09-23, the page is named next to the number, and where a vendor bills in something other than seconds we say so rather than quietly converting it.

The table everyone wants, with the unit printed next to every number

All prices are US dollars, list price, standard (not batch or enterprise) tier. "8-second clip" is simply the per-second rate times eight, so you can compare like for like; it ignores failed or rejected takes, which we come back to below.

Two notes on how to read it. First, Seedance is not billed per second. BytePlus bills per million tokens, and the page gives the formula: token consumption is roughly input plus output duration, times width, times height, times frame rate, divided by 1024. The per-second figures in the table are BytePlus's own worked examples for a 5-second, 16:9 clip with no video input, so they move if you change aspect ratio or feed in a reference video. Second, the Seedance 2.0 Mini row is list price, as is Seedance 2.0 Fast, which we left out of the table; the same page advertises a limited-time enterprise-only discount running until October 7, 2026, which we left out on purpose because most readers will not get it.

Runway resells third-party models through the same credit system. Its docs list Veo 3.1 at 40 credits per second with audio and 20 without, which is $0.40 and $0.20, matching Google's own Standard rate, and Seedance 2.5 at 20 to 68 credits per second depending on resolution. The model router is convenient; it is not automatically cheaper.

The widest gap is inside one vendor, not between them

Read the table vertically and something odd shows up. Within Google alone, the same 8-second clip runs from $0.40 (Lite, 720p) to $4.80 (Standard, 4k), a twelve-fold spread. Within Kling 3.0, 4K costs almost four times what 1080p does. Within Seedance 2.0, 4K costs more than ten times what 480p does.

Across vendors at the same tier, the numbers crowd together. Kling 2.5 Turbo at 720p, Veo 3.1 Lite at 720p, Seedance 2.0 Mini at 480p and MiniMax-H3-Max at 480p all sit between four and five cents a second. That is not a coincidence you should build a strategy on, but it does tell you where the real decision is:

• Resolution. If the clip ends up as a vertical short on a phone, paying for 4k is paying for pixels nobody sees.
• Native audio. Google's Vertex AI pricing page lists video-only Veo 3.1 at $0.20 against $0.40 with audio. If you add a voice track and music afterwards anyway, audio generation is a line you can often delete.
• Reference video. Kling and Seedance both charge more when you feed in a video, and Seedance applies a minimum token charge in that case.

Choose the resolution and the audio setting first. The brand name moves your bill less than those two switches do.

The number that is not on any pricing page

A per-second rate tells you what one generation costs. It does not tell you what one usable clip costs, and the two can be far apart. If one model gives you something you would post on the first try and another needs three attempts, the "cheaper" one is the expensive one. No vendor publishes that ratio because it depends on your prompts and your standards.

What the vendors do publish is who pays for failures. Google's pricing page says you are only charged if the video is successfully generated, and BytePlus says the same about Seedance, naming content moderation as one of the failure reasons that is not billed. That protects you from errors. It does not protect you from a clip that generated fine and is simply not good enough, and that is where most of the waste lives.

So the honest way to compare is a small test you run yourself:

1. Write ten prompts that look like your real work, not showcase prompts.
2. Run each on two or three candidate models at the resolution you will actually publish.
3. Count how many outputs you would ship without re-rolling.
4. Divide the total spend by that count. That is your cost per usable clip, and it is the only number worth putting in a spreadsheet.

Why the cheapest row today may not exist next year

The Sora shutdown is the obvious warning, and we covered what it does to cost per video in Sora 2 leaves the API and the replacement column is empty. It is not the only one. Google's own Gemini API deprecations page shows veo-3.0-generate-001, veo-3.0-fast-generate-001 and veo-2.0-generate-001 with a shutdown date of June 30, 2026, each pointing to a Veo 3.1 model. Kling's pricing page today lists 3.0, O1, 2.6 and 2.5 Turbo, and does not show a 2.1 row even though people still search for its price.

A commenter on Hacker News put the cost side of this in one line in June: "All of them are increasing prices and deprecating cheap models." (mistic92, 2026-06-26.) That is one developer's observation, not a measured trend, but the deprecation tables above are consistent with it. The practical consequence:

• Keep the model name and resolution in configuration, not in code.
• Re-read the pricing page on the day you change your budget, not the day you first chose a vendor.
• Keep at least one route that does not depend on a hosted generator at all, such as stock footage or your own clips, so a deprecation changes which route runs rather than whether anything runs.

That last point is how our own video side is built: NoobClaw's video workflows include a per-second generative engine alongside local remix, stock footage and template routes, with pricing shown in the app. It does not make any single model immortal. It means a model disappearing is a settings change instead of an outage. For the per-vendor detail, see our breakdowns of Veo 3 API cost per video and Kling API pricing by model version, and for creators rather than developers, Sora 2 alternatives picked by what you used it for.

FAQ: AI video generator API pricing

What is the cheapest AI video generator API right now?

At list price on 2026-09-23, the lowest per-second rates we found on official pages were around four to five cents a second: Veo 3.1 Lite at 720p ($0.05), Kling 2.5 Turbo at 720p ($0.042), Seedance 2.0 Mini at 480p (about $0.04) and MiniMax-H3-Max at 480p ($0.05). Which one is cheapest for you depends on how many takes you discard, so run the ten-prompt test above before committing.

Why do some vendors price per second and others per token or credit?

Per-second pricing is simple to read. Token pricing, as BytePlus uses for Seedance, ties cost to pixels and frames, so a square or lower-resolution clip genuinely costs less. Credits, as Runway and Kling use, let one balance cover many models. None is inherently cheaper; convert each to the cost of one finished clip at your real settings. We explain the metering problem in more depth in what you are actually buying from a video generation API.

Are these prices guaranteed?

No. They are what each vendor's page showed on 2026-09-23. Enterprise discounts, batch tiers and regional pricing are not included, and the deprecation tables quoted above show that whole models can disappear with a few months' notice. Check the vendor page again the day you sign anything.