Lyria 3.5 is single-turn and varies per call: keep the artifact

Google's Gemini API music page says music generation is a single-turn process, that iterative editing is not supported, and that results may vary between calls even with the same prompt. For Sume Music, which resolves sume/music-auto to Lyria 3.5 today, that means a track you like cannot be reproduced by re-sending the prompt. Download and store the artifact the first time.

What Google's page says

Model ids and the 30-second versus full-length split are from the same page.

What this means on Sume

Each Sume generation is a separate paid job at $0.125 per accepted generation, and job.request.routed_model records which engine ran. The result is an audio artifact on media.sume.com plus result.lyrics when present.

Re-rolling is therefore a deliberate cost. Generate a few variants in one batch with different Idempotency-Key values, pick one, and keep its job id.

import os, requests

for n in range(3):
    r = requests.post(
        "https://api.sume.com/v1/music-router/generate",
        headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}",
                 "Idempotency-Key": f"variant-{n}"},
        json={"prompt": "A 30-second upbeat marimba jingle. Instrumental."},
    )
    print(n, r.status_code, r.json().get("id"))

Editing without re-rolling

Trimming and joining do not need the model: use timeline audio to split a good section out or concatenate two takes.

Worked example

Two habits keep you out of trouble.

• Generate three variants for a key spot ($0.375 at $0.125 each) and choose one.
• Download the winning file and note its job id and routed_model.
• Fix small problems with timeline audio split or concat instead of re-rolling.

Checklist before you commit

Because Google says results vary between calls, do not use the same prompt as a stand-in for a saved file.

• Store the artifact in your own bucket.
• Write down the prompt.
• Re-check the Google page; the preview model ids may change.