How the price of a minute is calculated
No model generates a minute in one piece. They all work in segments. The reason is technical: to keep the video from falling apart, every frame must agree with all the others, and the difficulty of that grows faster than the length of the clip. So a minute is assembled from several scenes and stitched together in an editor.
Hence the formula:
Price of a minute = price of one segment × number of segments × regeneration factor
The first two factors are predictable and we calculate them below. The third is what nobody writes about, and it is what decides the real budget.
Table: one minute of video
Each model is counted at its maximum segment length, because that is cheaper: fewer segments means fewer seams and fewer regenerations. KADR prices, October 2026, sound on, base token pack (1 token = 1.58 rubles). On the large pack every figure below drops by about 18 percent.
| Model | Resolution | Max length | Clips per minute | Minute, tokens | Minute, rub. |
|---|---|---|---|---|---|
| Grok Imagine Video | 720p | 10 s | 6 | 480 | 759 |
| Google Veo 3.1 Lite | 1080p | 8 s | 7.5 | 532 | 841 |
| MiniMax Hailuo 2.3 | 1080p | 6 s | 10 | 550 | 870 |
| Gemini Omni Flash | 720p | 10 s | 6 | 660 | 1043 |
| Kling Video O1 | 720p | 10 s | 6 | 744 | 1176 |
| Alibaba Wan 2.6 | 1080p | 10 s | 6 | 792 | 1252 |
| Google Veo 3.1 Fast | 1080p | 8 s | 7.5 | 795 | 1257 |
| Seedance 1.5 Pro | 1080p | 12 s | 5 | 855 | 1352 |
| Kling O3 | 1080p | 15 s | 4 | 940 | 1486 |
| Kling v3.0 | 1080p | 15 s | 4 | 1052 | 1663 |
| Seedance 2.0 Fast | 720p | 15 s | 4 | 1084 | 1714 |
| Alibaba Wan 2.7 | 1080p | 10 s | 6 | 1242 | 1964 |
| Google Veo 3.1 | 1080p | 8 s | 7.5 | 2640 | 4174 |
| Seedance 2.0 | 1080p | 15 s | 4 | 3616 | 5717 |
| Seedance 2.5 | 1080p | 30 s | 2 | 5366 | 8484 |
Grok Imagine Video, Gemini Omni Flash, Kling Video O1 and Seedance 2.0 Fast are counted at 720p: they do not go higher. Gemini Omni Flash always generates with sound and holds up to 10 seconds. Kling Video O1 does only 5 or 10 seconds, so it is counted in 10-second segments. Hailuo 2.3 at 1080p generates only 6-second segments. Seedance 2.5 includes sound in its price. Kling Motion Control and OmniHuman 1.5 are not in the table: they do not invent a scene from scratch but work from your own material (Motion Control makes the person in a photo repeat the movements from your video, OmniHuman makes a photo speak or sing to your audio), and they are billed per second: about 13.4 and 19.5 tokens per second respectively. Sora 2 Pro is no longer in the table: OpenAI shut Sora down and switched off the API on 24 September 2026 (what to use instead). This is not a public offer. More on the most expensive model in the table in the Seedance 2.5 review.
The spread is about tenfold. A minute on Seedance 2.5 costs as much as about ten minutes on Veo 3.1 Lite. First conclusion: choosing the model saves more than all the other tricks combined.
Look at the clips column too. Models with long segments (Kling and Seedance 2.0 hold up to 15 seconds, Seedance 2.5 up to 30) need several times fewer generations per minute than Veo with its eight. That means not just money but fewer seams, and so fewer chances for a character to drift between scenes.
Resolution changes the price fivefold
The second conclusion is less obvious. Take one model, Seedance 2.0, and the same 6-second clip:
| Resolution | Tokens | Price, rub. |
|---|---|---|
| 480p | 69 | 109 |
| 1080p | 362 | 572 |
More than five times the difference for nothing. And while you are still tuning the prompt and do not know what you will get, you do not need resolution at all. You need to see whether the model understood the scene, whether the camera goes where it should, whether the character falls apart. All of that is perfectly visible at 480p.
Hence the rule that saves the most: drafts at 480p, the final at 1080p. Ten drafts and one final take on Seedance 2.0 come to 1090 plus 572, about 1660 rubles. The same ten attempts straight at 1080p would cost 5720 rubles.
A caveat: not every model has low resolution. Only the Seedance line does 480p. Kling, Grok, Wan and the junior Veo versions start at 720p, Hailuo at 768p, and the flagship Veo 3.1 exists only at 1080p, so you cannot make it cheaper with resolution.
Sound doubles the price
The third lever, which almost everyone forgets.
On most models with native audio, the «with sound» toggle nearly doubles the price. The exception is Seedance 2.5, where sound does not change the price. Here it is for a 6-second clip at 1080p:
| Model | Without sound | With sound |
|---|---|---|
| Google Veo 3.1 | 132 | 264 |
| Kling v3.0 | 71 | 106 |
| Seedance 1.5 Pro | 46 | 86 |
| Google Veo 3.1 Fast | 64 | 80 |
| Google Veo 3.1 Lite | 32 | 53 |
The question worth asking yourself: do you need sound from the model at all? If you lay music and voice-over in editing anyway, you pay double for a track you will throw away. Native sound is justified where lip sync and in-frame sound effects matter. Otherwise switch it off.
The hidden cost nobody mentions
Now the third factor. Every price above is the cost of a single generation. But you are not paying for a generation, you are paying for a usable result, and those are different things.
A simple static scene (landscape, macro, an object on a background) often works on the first or second try. A scene with a person talking and gesturing while the camera moves can take ten attempts or more: the face drifts, the hands do something impossible, the lips do not match the sound.
The honest way to budget: take the price of a minute from the table and multiply by 3 to 5. That is the usual range of attempts on hard scenes. A minute on Kling v3.0 turns from 1663 rubles into 5000 to 8300 rubles, which is a completely different conversation.
We deliberately do not give an «average» regeneration figure here: it depends on the prompt so much that any average would be a lie. It is better to set a factor and refine it on your first ten clips.
Compared with filming: when each one wins
A head-on comparison «generation versus production» makes no sense, because you are comparing different things. Compare by task.
Generation wins when a scene is expensive or impossible to film: shooting in another country, a historical period, a crowd, a special effect, an object that does not exist, a dangerous stunt. It also wins on time: what takes weeks of approvals in production takes an evening here.
Filming wins when the frame must show a specific real product, a specific person, or when you need legally clean footage with signed releases. No model will show your real product exactly as it is.
The practical takeaway for business: generation mostly does not replace filming, it covers what nobody used to film because of the budget. Nobody ordered a background clip for stories from a production studio, because the money made no sense. Now it costs about 840 rubles.
How to pay less for the same result
- Tune the prompt on a cheap model. Scene logic, composition and camera movement are worked out on Veo 3.1 Lite for about 84 rubles. Moving to Veo 3.1 or Seedance 2.5 makes sense only with a finished, debugged prompt.
- Drafts at 480p if the model supports it. The difference is fivefold, and you do not need resolution to test an idea.
- Keep sound off until you need it. Paying double for a track you will replace with music is pure loss.
- Start with short clips. The price grows linearly with duration: if a scene does not work at 6 seconds, it will not get better at 15, only more expensive.
- Use a model with long segments for long scenes. Kling and Seedance 2.0 hold up to 15 seconds, so a minute takes 4 generations instead of 7.5 on Veo, and Seedance 2.5 needs only 2. Fewer seams, fewer chances for the character to drift.
- Do not buy a subscription if you generate irregularly. In subscription services unused credits expire at the end of the month. With pay-per-use you spend exactly what you generated.
- Count the price of the result, not of the generation. A model that costs twice as much but gets it right the first time is cheaper than one that needs five attempts.
Image prices are collected in the article on how much it costs to generate an AI image, and a 30-second clip across different models is calculated here.
Bottom line
A minute of AI video in 2026 costs from 841 to 8484 rubles depending on the model, and that is without regenerations. The real budget comes from multiplying by 3 to 5.
The main levers, in order of impact: the choice of model (about a tenfold spread), resolution at the draft stage (fivefold), sound (up to twofold) and segment length. A subscription with irregular use is the most expensive option of all, because you pay for credits that will expire.
