A Reel is decided in the first second, muted, at phone size, by someone whose thumb is already moving. Everything about making one follows from that: vertical from the start, the visual hook in frame one, captions from the beginning, and enough variants that you find out which opening actually stops the scroll.
TL;DR
- Generate 9:16 natively. Cropping from 16:9 throws away 70 percent of the frame and costs the same
- Free tiers watermark, which is fatal here since a watermark reads as low effort
- Cheapest route: Wan 3.0 at 480p, 24 credits per 5 seconds
- With audio: Veo 3.1 Lite at 54 credits per 8 seconds. Ten variants is 540 credits
- First frame is the whole job. Compose the hook into frame one, not second three
Vertical from the start
The most common expensive mistake: generating 16:9 and cropping. You lose roughly 70 percent of the horizontal frame, your composition moves off centre, and you upscale a tall slice back to full resolution which softens it.
Ratio costs nothing. A 9:16 clip is the same credit price as 16:9 on every model. There is no saving to justify the loss. More in the AI video aspect ratio guide.
Composing for the phone
The hook in frame one. Not a build-up. The single most impactful change to a Reel is moving the interesting thing to the first frame.
Subject in the upper two thirds. The lower portion carries captions and platform UI.
Visual quiet in the lower middle third for text. Prompt for it: state that the lower part of the frame should be uncluttered.
Nothing important in the outer edges. Interface elements sit on the sides.
No text in the generated frame. It garbles once it moves. Add type in the edit where you control position per platform.
What it costs
Canon 8frame prices, a credit is $0.01 at pack rate:
| Route | Clip | Credits |
|---|---|---|
| Wan 3.0 at 480p | 5s | 24 |
| Wan 3.0 Prime, any resolution | 5s | 34 |
| Veo 3.1 Lite, silent | 8s | 33 |
| Veo 3.1 Lite, with audio | 8s | 54 |
| Kling v3 Standard, silent | 5s | 57 |
| Gemini Omni Flash, with audio | 8s | 135 |
| Music bed, 30s | Lyria 2 | 13 |
| Assembly | Merge Videos | 10 |
Ten hook variants on Veo 3.1 Lite with audio: 540 credits, about $5.40. That is the spend that actually matters, because you cannot predict which opening works.
Why variants, not polish
Reels performance is dominated by the first second, and nobody predicts it reliably. Ten different openings of the same idea, tested, beats one perfected opening by a wide margin.
Vary: the first frame, the first spoken words, the pace of the first cut, the caption's opening line. One change per variant so you learn something.
Speed matters here
Reels is a volume format and iteration speed is the constraint. On our own production jobs, Gemini Omni Flash has a 39-second median generation time against Kling's 143 seconds (70 and 58 jobs respectively, modest samples). When you are testing ten openings, that difference decides how many you actually try. See how long AI video generation takes.
What not to generate
Your face. If you are the channel, film yourself. The parasocial relationship is the asset and a synthetic version dissolves it.
Your product. Real product footage. Generated products misrepresent what arrives.
Anything with visible hands doing something specific. Reliable failure, and in vertical close framing it is unmissable.
Trending audio replacements. If a trend depends on a specific sound, use that sound, not a generated approximation.
FAQ
Does vertical cost more to generate? No. Same credit price as any other ratio.
What resolution for Reels? 720p is enough for most placements. Generate at delivery resolution, since there is no video upscaling on the canvas.
How many variants should I test? Ten openings of the same idea, varying one element each.
What is the cheapest way to make a Reel? Wan 3.0 at 480p, 24 credits, with your own audio over it.
Vertical from frame one, hook in frame one, ten variants. The 8frame canvas is free and unlimited, and generation is paid from $19/month.