What changed over 2026
Three shifts that were not there at the start of the year now define the choice of model.
Length. Seedance 2.5 (31 July), Wan 3.0 (generally available on 24 August) and the announced Kling 4.0 deliver half a minute of video in one pass. For a vertical video that means a scene assembled without stitching three pieces together.
Audio in one pass became the norm: Veo 3.1, Kling 3.0, Seedance 2.5, Wan 3.0 and MiniMax H3 all output speech, noise and music together with the picture.
Sora left the market: the app went offline on 26 April 2026 and the API on 24 September. How to rebuild production without it is covered in the piece "Replacing Sora".
What to look at when choosing models for business
The criteria below matter for quoting a batch, not for a pretty demo.
- Length of one clip: the fewer joins, the less that breaks at the seams.
- Native audio and language: not every model has Russian in native voiceover, otherwise you need separate speech synthesis.
- Resolution: 720p for drafts, 1080p for the client, 4K for image videos.
- References: they are the only way to keep a product or a face consistent between scenes.
- The price per second of finished video is the single number a quote is built from.
- Rights and watermarks: free tiers often ban commercial use.
Comparison table
Data as of 15 October 2026. Prices are official API rates in USD; aggregators show different numbers.
| Model | Clip length | Audio | Resolution | Price per second | Best for |
|---|---|---|---|---|---|
| Kling 3.0 (Kuaishou) | 3-15 s, multi-shot | yes | up to 4K | $0.08-0.17, 4K at $0.40 | volume of short videos |
| Kling 4.0 | up to 30 s (claimed) | yes | 10-bit HDR | early access since 28.09 | tests, full launch in October |
| Veo 3.1 (Google) | 4/6/8 s | yes | up to 4K, 9:16 | Lite $0.05, Fast $0.10-0.30, Standard $0.40-0.60 | on-screen text, advertising |
| Seedance 2.5 (ByteDance) | up to 30 s | yes | up to 1080p | via Runway API $0.20-0.70 | long scenes, up to 50 references |
| Wan 3.0 (Alibaba) | up to 30 s | yes | up to 1080p | $0.05 / 0.10 / 0.20 | scenes from documents and links |
| MiniMax H3 | up to 15 s | stereo | 768p / 2K | $0.08 / 0.13 | drafts, open weights |
| Vidu Q4 Preview | no data | yes | 540p-4K | $0.05-0.40 | cheap runs, since 7 October |
| Runway Gen-4.5 | 5-10 s | yes | upscale to 4K | about $0.12 | polishing filmed material |
Kling 3.0 and 4.0: volume at a low price
Kling 3.0 launched in February 2026 and remains the main model for batches: clips up to 15 seconds, storyboarding with shot changes inside one generation, native audio in five languages.
On the API price list a second costs $0.08 at 720p without audio, $0.11 at 1080p without audio, $0.13 and 0.17 with audio, and $0.40 in native 4K. A hundred ten-second videos in Full HD with audio comes to about $170 of generation.
Kuaishou announced Kling 4.0 on 30 September: native 30 seconds, 10-bit HDR, multiple keyframes. Early access is handed out by subscription and the full launch is set for October, so it is too early to build a stream on it.
The weak spot: Russian is not on the list of native voiceover languages, so it is done with a separate synthesis.
Veo 3.1: on-screen text and native vertical
Google released Veo 3.1 in October 2025, and in the Gemini API it is still the main video model: clips of 4, 6 and 8 seconds, 9:16 without cropping, stereo audio in one pass.
The main argument is readable text: a label, a sign or a print on packaging holds up better than with most competitors.
Gemini API prices with audio: Lite $0.05 at 720p and $0.08 at 1080p, Fast $0.10 / 0.12 / 0.30 at 720p / 1080p / 4K, Standard $0.40 and $0.60 in 4K.
Since May 2026 video in the Gemini app is generated by the new Gemini Omni Flash model; its API access is still expanding, so production keeps using Veo 3.1 for now.
Seedance 2.5: long scenes and plenty of references
ByteDance released Seedance 2.5 on 31 July 2026, with the open API appearing on 7 August. The model generates up to 30 seconds at a time and takes up to 50 references: photos, video and audio.
For business this is the direct route to keeping one product or character in a long scene without manual stitching. It also does targeted edits: swap the product or the background while keeping the motion.
The downside is the price. Via the Runway API a second costs about $0.20 at 480p, $0.30 at 720p and $0.70 at 1080p; on ByteDance's own platform you pay in tokens.
Wan 3.0, MiniMax H3 and Vidu Q4: the cheap alternatives
Wan 3.0 from Alibaba has been generally available since 24 August: up to 30 seconds, first and last frame, references and even generation from a document, slides or a web page. A second costs $0.05 at 480p, $0.10 at 720p and $0.20 at 1080p.
MiniMax H3 came out on 31 July 2026 and got open weights a few days later: up to 15 seconds in 2K with stereo audio, $0.08 per second at 768p. It works well for rough motion passes.
Vidu Q4 Preview from Shengshu launched on 7 October: from $0.05 per second at 540p to $0.40 at 4K, with up to 15 references. The model is fresh, so run it in tests rather than straight into client batches.
Runway Gen-4.5: polishing, not generating from scratch
Runway is valuable for what it can do with existing material: masks, repainting a clip through Aleph, retiming, character-reference work. Gen-4.5 gives 5-10 second clips with audio at about $0.12 per second.
Since September Runway has a Premiere Pro plugin with Gen-4.5, Seedance 2.5, Kling 3.0 Pro and Veo 3.1 inside it, so models can be compared right on the editing timeline.
What a batch costs: the maths
The job: 100 vertical videos of 15 seconds each in 1080p with audio, generation only.
Kling 3.0: 1,500 seconds × $0.17 ≈ $250. Wan 3.0: 1,500 × $0.20 ≈ $300, and the scene comes out as one clip. Veo 3.1 Fast: 1,500 × $0.12 ≈ $180, but 15 seconds has to be built from two generations.
On top come rejects, voiceover, music, subtitles and adaptation to three platforms. The price per second is the input to a quote, not the cost of a video.
The short version
For a stream of vertical videos: Kling 3.0 is the main model on price, Veo 3.1 for on-screen text, the ByteDance and Alibaba models when a long scene with one character is needed, Runway for polishing footage. MiniMax H3 and Shengshu's Q4 work for drafts, and Kling 4.0 is worth checking after the full launch.
What next
How to build a stream of videos out of all this: a walkthrough of the pipeline from brief to publishing. How to prepare copies for geos and platforms: the article on versioning. Production in batches from 100 videos a month: videos made with AI and the mass vertical videos format.