Best of 8 music takes costs $1.00 on Sume, then a 1-cent cut

Eight music takes cost $1.00 on Sume, because Music 1.0 and the Music Router charge a flat $0.125 for each accepted generation, whatever the prompt length. Cutting the winner to length with a Timeline audio split costs another $0.01, so $1.01 buys you a choice of eight. There is no duration parameter: the Music docs say Sume rejects duration and duration_seconds.

Why generate several

The Music docs tell you to treat the seven brief axes as creative directions, not guaranteed output, and to examine the generated audio. They also say the provider's lyrics field is model-reported metadata and not an audio measurement. So the plan is to write a precise brief, generate a handful, listen, and keep one. Because the price is fixed per generation and does not change with prompt length or image conditioning, the cost of a choice is simply the number of takes times $0.125.

Steering the length without a duration field

Put the length in the prompt, such as a 30-second track, or use section markers like [0:00-0:30] Intro. Tracks can still come back longer or shorter than asked, so the cut is your safety net. The router accepts the same body as Music 1.0 and adds an optional model: sume/music-auto by default, or lyria-3.5 or lyria-3-pro pinned. Every router model charges the same fixed music price.

Cutting the winner

Timeline audio takes operation split, a Sume-hosted url and up to 20 ranges, each with a start and an optional end. Cut the 30 seconds you want with one range, or several hooks with several ranges, and still pay one $0.01 job. Use WAV output if you will join the cut again; MP3 adds padding at each edge.

The music is delivered as a Sume-hosted audio artifact, so it satisfies the same-workspace rule for Timeline audio with no import step. Do not send exclusions as negative_prompt: a non-empty value returns HTTP 400. Write them into the positive prompt instead, ending with Instrumental, no vocals when you want no singing.

Keeping takes organized

Name each generation with the brief you used, store its job id with your project notes, and score the takes against the brief in a short list. A blind listen in random order is cheaper than a second round of eight, which would be another $1.00. If none of the first eight fits, change one axis of the brief and run three more, $0.375, rather than rewriting the whole prompt. The router lets you pin lyria-3.5 or lyria-3-pro for a comparison, at the same price.

What eight takes tell you

If three of eight are usable, the brief is working. If only one is usable, the brief is probably too loose, and another eight with the same words will not help; tighten one axis and rerun. If none are usable, check that you did not send a duration or a negative prompt, since those are rejected, and that exclusions are written in the positive prompt instead.