Which AI video API should you use? Pick by task

Pick by task

Prices are US dollars per second of video at the setting shown, from the public catalogue. Each model's page lists exactly what it accepts.

These are starting points chosen from what each model accepts and what it costs, not a ranking of quality. Quality depends on your prompt and your footage, which is why the last section is about testing.

Decide in five questions

1. What do you have? Only text, a still image, a video, or a voice track? A lip-sync model needs an image and audio. An edit model needs a video. A plain text to video model needs neither.
2. Does it need sound? Some models make audio in the same pass and some sell it as an option. Veo 3.1 charges $0.20 per second without audio and $0.40 with it at 720p.
3. How long is a clip? Check the duration range. Seedance 2.5 takes 4 to 30 seconds and Veo 3.1 takes 4, 6 or 8 seconds.
4. Which resolution is the final? Draft at 480p or 720p and render the final at 1080p. 4K is not worth paying for in a draft.
5. What is the budget per usable clip? Multiply the per-second price by the length, then by how many tries you expect. The price comparison guide shows how.

Test two models on the same prompt

The cheapest way to choose is to run the same prompt on two models at 480p or 720p and look. A 5 second 720p test is $0.15 on Wan 2.6 Flash and $1.05 on Seedance 2.5. With the API that is two calls that differ in one field:

for model in wan-2.6-flash seedance-2.5; do
  curl -s -X POST https://vidlab.ai/v1/generations \
    -H "Authorization: Bearer $VIDLAB_API_KEY" \
    -H "Content-Type: application/json" \
    -H "Idempotency-Key: $(uuidgen)" \
    -d "{\"model\":\"$model\",\"input\":{\"prompt\":\"A barista pours milk into a latte, close up\",\"resolution\":\"720p\",\"duration\":5,\"aspect_ratio\":\"9:16\"},\"max_price_usd\":1.10}"
done

A model refuses an option it does not accept instead of quietly changing it. Read the model's params from GET /v1/models/{id} and send only what it lists.

Why use one API for all of them

Video models change every few weeks. A new model that suits your task should not mean a new account, a new SDK and a new invoice. On VidLab:

• One key and one request shape work for all 137 models, across video, image, audio, lip sync and upscaling.
• Every call can be priced first, for free, with POST /v1/estimate.
• Each call goes to the cheapest route that is up, and the price you pay is the catalogue price.
• A failed generation is not charged.
• Agents can use it too, through MCP or the skill file.