How many AI music tracks should I generate for one ad before picking?

Generate three tracks for one ad, and pick the best. Every Music Router generation is a flat $0.125 on Sume, whatever the model, so three takes are $0.375. If the brief is tight (tempo, mood, length, instrumental), three is usually enough to find one that fits. If the brief is loose, a cap of five is a better rule. Rates were read on 2026-10-07.

This is a budget question, not a quality one. At twelve and a half cents, a take costs less than one second of 720p video at the Omni rate, so the cost of picking among tracks is small compared with the cost of the clips they sit under. The real cost is your time listening.

Why a flat price changes the rule

Because the price per generation is fixed, there is no saving in asking for a shorter track. The prompt controls the length, and a 15-second track and a 90-second track both cost one generation. So ask for the length you need, and put the structure in the prompt with times: [0:00-0:05] Build, [0:05-0:12] Main, [0:12-0:15] Sting.

A second consequence: do not stack effects in one prompt to save generations. One mood per prompt gives clearer takes, and you compare like with like.

A take budget

Use the brief to decide. If you already have a reference and can name tempo, key, instruments and length, generate three and pick. If you have a mood only ('warm, festive'), generate five, and use the best as the reference for a tighter second prompt. Past eight, change the prompt: more takes of the same prompt give you the same family of sounds.

Audition with the video, not on its own. Put each take under the cut in a draft render and listen with the picture. A Timeline 1.0 render is $0.10 per started minute, so three draft renders of a 30-second ad are $0.30. That is more than the three tracks, so audition the tracks first, and render only the top two.

curl -X POST https://api.sume.com/v1/music-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ad-track-take-1" \
  -d '{
    "model": "sume/music-auto",
    "prompt": "A 15-second upbeat holiday-sale track, 118 BPM, bright synth plucks, one clear drop at 0:05, ends on a short sting. Instrumental, no vocals."
  }'

Keep the takes you did not use

Every track is a durable file with an id. Name the takes in your notes by prompt and date. A losing track for a sale ad may be the right bed for a product demo next month, and you already paid for it.

Change only one thing between takes when you can. Use the same idempotency key only for a true retry of the same request: a new take is a new key, and a retry of a failed submit reuses the old one.

A quick way to listen

Listen to the first 5 seconds and the last 5 seconds of each take before the middle. The start tells you whether the track hooks, and the end tells you whether it finishes or just stops. A track that ends badly is not a failure if you can end the ad earlier: trim the audio to the length of the cut with a Timeline audio split range ($0.01), and use a fade in the render.

Write one line per take: the idea, the first thing you liked, the first thing you did not. After three takes, the notes show what to change in the prompt, and the next take is better aimed than the first.

If you are making a family of ads, pick one track and cut it three ways, rather than three different tracks. A single sound across the set is a brand.

Cost by number of takes

• The music cost grows by $0.125 a take, and a draft render adds $0.10 a time.
• The final render and captions are the same whichever track you pick.
• If the same prompt gives three sounds you dislike, rewrite the prompt before asking for more.