Tools

AI Video Model Prices Compared: Veo 3.1, Kling 3.0 and Seedance 2.0 After the Sora API Shutdown

Per-second prices now run from $0.03 to $0.778. Here is what an 8-second clip really costs on each major video API, and where the hidden multipliers sit.

NP

Nolan Pierce

Tools & Platforms Editor

Published Oct 27, 2026 7 min read

AI video API prices in October 2026 run from $0.03 to $0.778 per second, and the median 8-second clip bought directly from a model maker costs $0.93. Veo 3.1 Lite and Fast sit at the low end, Kling 3.0 in the middle, and Seedance 2.0 at 4K at the top. OpenAI's Sora API closed on September 24.

This comparison is for developers, producers and product teams who need to budget a video feature and are choosing between APIs. It covers per-second prices, what an 8-second clip costs at common settings, what changed when Sora left, and how audio, resolution and marketplaces change the bill. Model Drop tracks new models, platforms and the tools built on them, and video is where list prices differ the most from what a finished clip costs.

What Does AI Video Cost per Second Right Now?

The most complete snapshot comes from UGC Copilot, a video tools vendor that read list prices from each provider's pricing pages on October 1, 2026. It covers 68 price points across 18 models from eight companies, and it uses 48 direct-from-maker per-second prices for its median, according to its price index.

The range is wide. The cheapest price is $0.03 per second, for Veo 3.1 Lite on Vertex AI at 720p without audio. The most expensive is $0.778 per second, for Seedance 2.0 on BytePlus at 4K with audio. Of the 48 direct prices, 27 cost under $1.00 for an 8-second clip.

The index uses list prices before tax, discounts, promotions or subscriptions. Treat it as a ceiling for planning, and confirm current rates on each vendor's page before you commit a budget.

How Do the Main Video Models Compare on Price?

The figures come from the UGC Copilot index. Google's own pricing page confirms one related number: Gemini Omni Flash works out to about $0.10 per second of 720p video, based on 5,792 output tokens per second, according to Google's Gemini API pricing page. The Veo rates were not on that page when we checked, so confirm them in the Veo documentation.

What Does an 8-Second Clip Actually Cost?

Multiply the per-second rate by eight. That math ignores retries and input fees, but it gives a floor. Here are 1080p examples from the index:

The right column is the useful one. A team that needs 1,000 finished clips a month spends under $1,000 on the cheap tier and nearly $3,000 on the premium one. If you need three takes to get one usable clip, multiply each figure by three.

What Did the Sora API Shutdown Change?

It removed one option and pushed buyers toward Google, Kling and ByteDance. OpenAI shut down the Sora web app on April 26, 2026 and the API on September 24, 2026, as part of a shift toward coding tools and enterprise customers, according to The Decoder. Sora continues as a research project on world models.

UGC Copilot notes that OpenAI's model page lists "no one-to-one replacement API." Sora 2's last published price was $0.10 per second at 720p with audio, and Sora 2 Pro ran from $0.30 to $0.70 per second. The nearest live 720p models with native audio are Veo 3.1 Fast at $0.10 and Gemini Omni Flash at about $0.10.

If your product called the Sora API, the migration is not a drop-in swap. Prompt styles, clip lengths, aspect ratios and audio behavior all differ, so budget time for re-testing, not only a new price.

Does Audio Change the Price?

Often it does. Kling 3.0 charges $0.112 per second at 1080p without audio and $0.168 with it, a 50% premium. Veo 3.1 on Vertex AI, with audio off, lists at $0.20 per second at 1080p, while the Gemini API with audio on lists at $0.40. The two are not an exact like-for-like, since the platforms differ, but the gap shows audio's weight.

Native audio can replace a separate sound step, which may save time and money. If you plan to add your own music and voice anyway, switch audio off and save the premium. Our look at text-to-speech models compared on latency, naturalness and cost shows what a separate voice track costs.

Why Does 4K Cost So Much More?

Because the model does more work per frame. UGC Copilot found 4K costs 50% to 150% more per second than 1080p across seven official price pairs. The smallest gap was Veo 3.1 on the Gemini API, from $0.40 to $0.60. The largest, 150%, applied to Veo 3.1 Fast and Kling 3.0.

For most uses, 4K is wasted at the draft stage. Generate at 720p or 1080p, pick the best takes and upscale or regenerate only the finalists. A social clip viewed on a phone rarely needs a 4K master.

The same logic shows up in images, where higher resolution drives cost. We covered that in our AI image model price comparison of Nano Banana 2.1 and FLUX 3 Image.

Are Marketplaces Cheaper or More Expensive?

Often more expensive. Seedance 2.0, Seedance 2.0 Fast and Seedance 2.5 cost 82% to 114% more per second on fal.ai than on BytePlus in the index. Kling 3.0 cost the same on fal.ai as on Kling's official API, so the markup depends on the model.

Marketplaces earn their fee through convenience: one account, one invoice and one integration. That can be worth it for a small team. For a large workload, a 100% markup is a six-figure question, so run the direct price against the marketplace price every quarter. We explained why price is the wrong first question in our inference providers comparison.

What Does the Price Index Leave Out?

Several costs sit outside per-second list prices. Luma prices by the clip, not by the second, so its Ray3.2 rows have no per-second figure: a 5-second 540p clip lists at $0.15 and a 10-second 1080p clip at $3.60. Pika 2.5 lists text-to-video at $0.04 per second at 720p, but requires a $10 monthly membership, which changes the math for low-volume users.

The index also ignores single-clip length limits and input fees. If a model caps clips at five seconds, a 30-second scene takes six calls, and each call can carry its own setup cost. Image-to-video workflows may add the price of generating the starting image.

Finally, list prices are not what enterprise buyers pay. Volume discounts and committed-use deals can change the ranking, so treat the table as the opening bid in a negotiation, not the final number.

Which Video Model Fits Which Job?

Match the model to the stage of work. Use low-cost options for drafts and testing, and reserve premium models for final output. A simple split:

Model Drop's editors have not run a quality test across all of these models, so this list reflects price and positioning, not output quality. Price tells you what a clip costs, not whether it is usable. Run your own prompts before you commit.

How Should You Budget a Video Feature?

Budget for takes, not clips. Follow these steps:

Prices in this market move quickly, so lock in flexibility. Prefer pay-as-you-go over annual commitments until the market settles, and write your integration so you can swap models without a rewrite.

Conclusion

AI video now costs from a few cents to about 78 cents per second, and the real bill depends on audio, resolution, retries and where you buy. Model Drop's editors compared the vendor index with Google's own page and found the same pattern: the sticker price is only the starting point.

Your next step is a one-week pilot. Draft 50 clips at 720p on two models, count the takes you need, and price the winner at your final resolution before you sign anything.