A four-scene score on Sume Music Router: tempos 24 BPM apart, $0.50

For a four-scene film that should feel like four scenes, send four separate Music Router jobs, change the genre family, the tempo and the lead instrument between them, and keep the tempos at least 12 BPM apart as the Music docs advise. Tempos of 72, 96, 120 and 144 BPM are 24 apart. Four jobs at $0.125 each cost $0.50, whatever the prompt length.

The rule the docs give

The Music 1.0 docs say that for scenes in one project that contrast, you change the broad genre family, the tempo (at least 12 BPM apart), and the lead instrument, and that if the user wants one consistent score you keep continuity instead. They also say the model has no duration parameter, so length is written into the prompt, and the axes in a brief are creative directions, not guaranteed output values.

The request

Send each brief to POST /v1/music-router/generate with model omitted or set to sume/music-auto. Add the length to the prompt text, such as "A 15-second track." Do not send duration or duration_seconds; Sume rejects both. Each scene needs its own Idempotency-Key so a retry does not buy a second track.

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: score-scene-3" \
  -d '{
    "model": "sume/music-auto",
    "prompt": "Drill, 120 BPM half-time pulse, E phrygian. 808 with long glide, muted hi-hats, a detuned bell line. A 15-second track. Tight and dry, 2024 production. Instrumental, no vocals."
  }'

When one consistent score is better

Contrast is a choice, not a rule for every film. If the four scenes are one story and the user wants a single mood, generate one longer track instead, and keep the genre, tempo and lead instrument the same across any follow-ups. The docs suggest passing the accepted scene still as image_url when continuity matters, so the music is conditioned on the same picture. Image URLs must be public HTTPS. The price does not change with the image: it is still $0.125 per accepted generation.

After a rejection

If a prompt is rejected by policy, the docs say to change the flagged content but keep the musical brief, and to retry only within your authorized budget. Do not reduce the request to a generic bed. The catalog says failed or pre-generation-canceled jobs are refunded, but a retry that is accepted is a new generation at $0.125, so a $0.50 score becomes $1.00 if every scene is retried once.

Check the result before you cut

Because length is steered by words, the returned file will not always be 15 seconds. Read the audio artifact from result.artifacts[] and measure it. A track that runs long can be cut with a Timeline audio split range, which is $0.01 per job, and job.request.routed_model tells you which engine ran. At $0.125 a track, a second pass on the one scene that missed costs less than rewriting the whole brief.