1. Drafts at low resolution
While you are checking the camera movement and composition, you do not need high resolution.
| Seedance 2.0, 6 seconds | Tokens | Rubles |
|---|---|---|
| 480p | 69 | ~109 |
| 1080p | 362 | ~572 |
Savings: 5 times. 10 drafts at 480p plus a final at 1080p cost about 1660 rubles, 10 attempts straight at 1080p about 5720.
2. Turn off the sound if music goes on top
| 6 seconds, 1080p | Without sound | With sound |
|---|---|---|
| Veo 3.1 | 132 | 264 |
| Seedance 1.5 Pro | 46 | 86 |
| Veo 3.1 Lite | 32 | 53 |
Figures in tokens. On Seedance 2.0 and 2.5 sound is included in the price, so turning it off saves no tokens there.
Savings: up to 2 times. For reels, music videos, ads with a voiceover and video covers you do not need the model’s own sound. More: how to add voice to a video with AI.
3. Test on a cheap model
Work out the prompt on Seedance 1.5 Pro at 480p without sound (10 tokens, about 16 rubles) or on Veo 3.1 Lite. Once the scene works, move the prompt to a flagship. Model comparison: which AI to choose for video.
4. Image first, then video
Debug the shot as an image in Nano Banana Pro (about 25 rubles), then animate it. Redoing an image is cheaper than redoing a clip. More: image to video or text to video.
5. Short clips
The longer the clip, the more expensive a failed attempt and the higher the chance that something breaks. For tests, take 5 or 6 seconds. Build a long video from short scenes: long video with AI.
6. Image quality levels
GPT-5.4 Image 2 has three levels: low 2 tokens, medium 6, high 17. Drafts on low, the final on high: more than an 8x difference. Prices of all models: how much it costs to generate an image.
7. Not the flagship out of habit
For a reel with a simple action, Veo 3.1 Lite (about 84 rubles for 6 seconds with sound at 1080p) looks almost the same on a phone screen as the senior Veo 3.1 (about 417 rubles). You need the flagship where the difference shows: close-ups, complex physics, a big screen.
8. A precise prompt the first time
Failed attempts are the biggest cost item. One action, one camera move, specific words. On KADR (kadr-ai.ru) the «Improve» button rewrites the prompt for free and suggests the model and settings: up to 10 times a day with a positive balance and up to 3 times with a zero balance. Templates: AI prompt generator.
9. Avoid what the model will refuse
A moderation refusal is lost time. Check real faces in Seedance, and celebrities and brands everywhere, in advance: the AI refuses to generate.
10. A larger token purchase
| Purchase amount | Token price |
|---|---|
| Packs of 100, 250, 500 and 1000 tokens | ~1.58 rub. |
| From 2000 tokens | ~1.49 rub. |
| From 10,000 tokens | ~1.40 rub. |
| From 50,000 tokens | ~1.30 rub. |
An amount larger than the ready packs is entered manually when you top up. KADR prices, October 2026. Tokens never expire, so a large purchase pays off even if you generate irregularly.
How much it all saves together
Example: an ad clip of 3 scenes, 3 attempts per scene.
| Approach | Calculation | Approximately |
|---|---|---|
| All attempts straight on Veo 3.1, 1080p, with sound | 9 × 264 tokens | ~3760 rub. |
| Drafts on Veo 3.1 Lite without sound, final on Veo 3.1 without sound | 6 × 32 + 3 × 132 tokens | ~930 rub. |
Savings: about 4 times with no loss of quality in the final. This is not a public offer.
In short
The biggest savings come from drafts at 480p (5 times), sound turned off (up to 2 times), tests on a cheap model, an image before the video and a precise prompt the first time. Use the flagship only for the final, a large token purchase is cheaper than small packs, and tokens on KADR never expire.
