A/B test two AI music beds on one ad: cost and the Timeline steps

Testing two music beds on the same ad costs less than a coffee when the ad is 30 seconds long. Each bed is one generation at a fixed $0.125 on Sume's Music Router, and each rendered cut on Timeline 1.0 is $0.10 per output minute, rounded up, so a 30-second ad rounds to one minute: $0.10. Two beds and two renders come to $0.45.

Cost for one test

How to keep the test fair

Change one thing. Keep the video slots, voice and audio.gain_db identical and change only the soundtrack.url. Set the same gain_db and fade_out_seconds on both. If the ad has a voiceover, set the same duck_db on both too, because ducking changes how loud the bed feels more than the track does.

Write two briefs that differ on one axis, for example tempo or instrumentation, rather than two unrelated prompts. For Google's reference, Lyria 3.5 is listed at $0.08 per song on its pricing page, read on 2026-10-05; Sume's price is the figure you pay.

Running it

Call POST /v1/timeline-1.0/plan first for each cut: it is an unbilled preflight and catches errors such as soundtrack_fade_exceeds_output before you spend a render. Then post both renders, using a distinct Idempotency-Key per variant so a retry does not double charge. Label the two outputs A and B in your own storage and run them as separate ad sets; the platform's own split-test tool does the measuring. Budget a little extra for a third take if neither bed lands: each regeneration is another $0.125, and the 30-second render for the winning bed is still only $0.10. Because music is a single generation with no edit pass, a bed you almost like is replaced rather than adjusted, so keep the prompts saved next to each result.