Castaly · Blog · Guides
Text to Video: How to Pick a Model, Write the Prompt and Control the Cost
Go from a sentence to a finished AI video clip: which models accept text alone and what they cost, a prompt structure that works, and settings that save money.
Text to video is the simplest way to start: write what should happen and get a clip back. It is also where most first attempts waste money, because a vague sentence leaves the model to make every decision and a default setting can cost ten times what the test needed.
This guide covers the three decisions that matter: which model, what to write, and which settings. Every prompt quoted is the exact text of a recorded generation.
Step 1: pick a model that takes text alone
Most video models in Castaly accept a prompt with nothing else. Four do not: wan2.7-i2v, wan2.7-r2v, happyhorse-1.1-i2v and happyhorse-1.1-r2v need an image or a reference.
A 5-second clip at 16:9 with audio off:
| Model | 720p | 1080p | Longest clip |
|---|---|---|---|
seedance-2.0-mini |
$0.19 | — | 15 s |
kling-2.6 |
$0.19 | $0.32 | 10 s |
seedance-2.0-fast |
$0.37 | — | 15 s |
wan2.7-t2v |
$0.37 | $0.56 | 15 s |
kling-3.0 |
$0.38 | $0.51 | 15 s |
wan3.0-video |
$0.49 | $0.99 | 30 s |
happyhorse-1.1-t2v |
$0.50 | $0.64 | 15 s |
seedance-2.0 |
$0.94 | $2.33 | 15 s |
seedance-2.5 |
$1.44 | $3.54 | 30 s |
For a first project, Kling 3.0 at 720p is a reasonable default: it is cheap, and it can go to 1080p or 4K when you find a shot worth keeping. If you need more than 15 seconds in one take, start on Wan 3.0, the cheaper of the two 30-second models. The video price comparison lists every model.
Step 2: write four things
A working text-to-video prompt answers four questions. Here is one from a 5-second Seedance 2.5 clip:
A black horse gallops through shallow sea water at sunset, spray exploding around its hooves, mane and tail streaming. Low tracking shot alongside, then the horse leaps and the camera stays low as water arcs through the backlight. Warm amber against deep teal, photoreal, powerful slow motion, no text.
| Question | In the prompt above |
|---|---|
| What happens? | A black horse gallops through shallow sea water at sunset, spray exploding around its hooves |
| Where is the camera? | Low tracking shot alongside, then the horse leaps and the camera stays low |
| What does it look like? | Warm amber against deep teal, photoreal, powerful slow motion |
| What should stay out? | no text |
Three habits make the difference:
- Describe action, not a picture. A video prompt needs a verb and a change. "A horse on a beach" gives the model nothing to do for five seconds.
- Size the event to the length. Five seconds holds about one action and its result.
- Say "no text". We end nearly all our prompts with it, because stray lettering is hard to remove afterwards.
The prompt guide takes the structure further, with seven parts for more ambitious shots.
The style goes in the prompt too
Text to video is not only for photoreal footage. Put the style first when it is the point of the clip:
Minimal black-ink line animation on a white page: a small cat chases a butterfly from left to right, the lines are loose hand-drawn sketch strokes that wobble slightly frame to frame, the butterfly loops once and the cat pounces and misses. Flat white background, no shading, no text.
A calm, wide shot needs only a few words about the camera:
A fleet of hot-air balloons drifts over mist-filled mountain valleys at dawn. Aerial camera glides slowly between the balloons as golden light breaks over the ridges and fog rolls through the pines below. Calm, majestic, photoreal landscape, steady motion, no text.
Step 3: choose settings in this order
- Ratio. Choose it for where the clip will play: 16:9 for a site or YouTube, 9:16 for TikTok, Reels and Shorts. It rarely changes the price, and you cannot crop your way from one to the other without losing the framing.
- Length. Start at 5 seconds. Length is the one setting that multiplies the bill directly.
- Resolution. Test at the lowest resolution the model offers, and render the final at the one you need.
- Audio. Leave it off unless you need the sound. On Kling 3.0 it adds 50%; on Seedance and Wan 3.0 it adds nothing.
A cheap way to work
Settle the prompt on a cheap model, then spend on the shot you keep. On Seedance 2.0 Mini, a 5-second 480p draft is $0.09, so a dozen variations cost about a dollar. When a prompt is working, render it on the model you want for the final.
The final is a new generation, so it will not match the draft frame for frame. Use drafts to check that the idea, the framing and the timing work.
When text is not enough
Text leaves the model free to choose every detail, and it chooses again on every run. Move to images when you need:
- The same character or product across clips. Use reference images.
- An exact composition. Make a still first and use it as a first frame.
- A specific ending. Supply a last frame.
The image-to-video guide shows all three.
FAQ
What is the cheapest text-to-video model on Castaly?
Seedance 2.0 Mini at 480p: about $0.09 for 5 seconds. At 720p, Seedance 2.0 Mini, Kling 2.6 and Kling 2.5 Turbo all cost about $0.19.
How long can a text-to-video clip be?
Up to 30 seconds on Seedance 2.5, Wan 3.0 and Wan 3.0 Prime. Most other models stop at 15 seconds, and Kling 2.6 and 2.5 Turbo at 10.
Should I write prompts in English?
All the prompts on this page are in English, and that is what our sample clips were made with. We have not tested other languages systematically.
Why did the same prompt give a different result?
Each generation is a new run, and the model makes its own choices wherever the prompt is silent. The more you specify, the less it varies. For a fixed look, use a first frame or reference image.