A/B test two music beds on one cut: $0.45, then $0.225 for each extra

Testing two music beds on the same cut costs $0.45 on Sume: two Music generations at $0.125 and two Timeline renders at $0.10 each, for a cut of 60 seconds or less. Every further bed adds $0.225. Five variants are $1.125, and the footage, which you pay for once, is not in these numbers.

The cost scales linearly with the number of variants, so the test is cheap compared with the footage. What is hard is the reading of results, and this post does not invent any: whichever platform you publish on will tell you what performed, and you should rely on its figures.

The cost of N variants

For a cut of up to 60 seconds, a variant is one bed ($0.125) plus one render ($0.10) = $0.225. A cut of 61 to 120 seconds bills two render minutes, so a variant is $0.125 + $0.20 = $0.325.

Make the variants differ

Two beds that sound alike test nothing. The Music docs recommend changing the broad genre family, the tempo (by at least 12 BPM) and the lead instrument when scenes should contrast. Apply the same rule to variants: for example 84 BPM lo-fi with Rhodes against 108 BPM synth-pop with a plucked lead. Keep the rest of the brief (length, Instrumental, no vocals) identical so the bed is the only variable.

Because the model takes no seed, you cannot regenerate the same bed. Keep the artifact URL of the winner; the Music docs say Sume returns hosted audio artifacts on media.sume.com.

Keeping the results traceable

Send metadata such as {"variant": "A", "cut": "spring-59"} with each Music request. The docs say Sume stores it on the job and does not send it to the provider, so it comes back when you read the job. Use a distinct Idempotency-Key per variant (for example spring-59-bed-A) so a retry does not create a second paid job.

Publish the variants under the same conditions, as separate uploads or as the platform's own split test allows. The Timeline soundtrack field takes gain_db, so also match the loudness of both beds before you compare them: a bed that is 6 dB louder will win a listening test for the wrong reason.

Cap the experiment before you start. Decide the number of variants first (two or three is enough for a single cut), write the briefs side by side, and generate them in one pass. Each Music job is a single $0.125 charge with no duration field to tune, so there is no way to overspend on one bed, but an open-ended loop of re-rolls adds up: ten tries on one cut is $1.25 before any render. Keep the winner and the artifact URL, discard the rest, and reuse the winning bed on the next cut in the series, where a render alone is $0.10.