Benefits of Lyria 3.5 for Campaign Music on a Budget

The campaign needs a track by Thursday, the licensing quote came back with a comma in an uncomfortable place, and the client wants three cuts. Lyria 3.5 campaign music is Google's answer for the drafting stage: lyria-3.5 went generally available in the Gemini API on September 3, 2026 (changelog). It makes full-length songs from text, or text plus images, at 44.1 kHz stereo.

Below are five benefits for marketing teams, plus the case where it isn't worth the trouble.

1. A flat $0.08 per track makes Lyria 3.5 campaign music easy to budget

The Gemini API pricing page lists lyria-3.5 at $0.08 per song on the paid tier. There is no Free tier and no batch discount listed. Your cost is the number of requests times eight cents.

Ten ad variants with five takes each is 50 songs, or $4.00. The same page says paid-tier content is not used to improve Google's products, which matters when your brief describes a product nobody has announced yet.

The catch: with no Free tier, you pay from the first test song.

import base64
from pathlib import Path

from google import genai

client = genai.Client()  # reads GEMINI_API_KEY from the environment

brief = (
    "Upbeat acoustic pop, 110 BPM, handclaps and warm guitar, "
    "feels like a sunny Saturday. Instrumental only, no vocals."
)

TAKES = 5  # 5 x $0.08 = $0.40
for take in range(1, TAKES + 1):
    interaction = client.interactions.create(model="lyria-3.5", input=brief)
    audio = interaction.output_audio
    if audio:
        Path(f"spring_sale_take{take}.mp3").write_bytes(base64.b64decode(audio.data))

This requests five takes of one brief and saves each MP3 under its own name. The cap is a loop count you can read in code review, so nobody finds the bill by surprise.

2. Section tags and timestamps give you control over song structure

The music generation guide says tags like [Verse], [Chorus] and [Bridge] help the model follow song structure. You can also write timestamp ranges, one segment per line, in the form [start - end] Label: description. Duration can be steered in the prompt too, for example by asking for a 2-minute song.

That is useful when the format is fixed. A 30-second social spot needs a hook in the first seconds and a clean ending. The docs example runs one minute, so shrinking it is my adaptation: test it before you promise anything.

prompt = """
[0:00 - 0:05] Intro: Bright plucked guitar hook, no drums yet.
[0:05 - 0:20] Verse: Handclaps and light drums join, warm and friendly.
[0:20 - 0:30] Outro: Build to one big hit, then a clean stop.
Instrumental only, no vocals.
"""

interaction = client.interactions.create(model="lyria-3.5", input=prompt)

The closing line asks for an instrumental, which is the easy way to get a bed that sits under a voice-over.

3. Image-led briefs let you start from the mood board you already have

Marketing briefs usually arrive as pictures, not music theory. lyria-3.5 accepts up to 10 images alongside the text, so the key visual can steer the track.

img_b64 = base64.b64encode(Path("desert_sunset.jpg").read_bytes()).decode("utf-8")

interaction = client.interactions.create(
    model="lyria-3.5",
    input=[
        {"type": "text", "text": "Ambient track inspired by this image's mood."},
        {"type": "image", "mime_type": "image/jpeg", "data": img_b64},
    ],
)

The input is a list of typed parts: one text part, then image parts carrying base64 data and a MIME type. Say in the text what you want from the picture, otherwise you are asking the model to guess.

4. Every track carries a SynthID watermark

All audio from Lyria 3.5 includes an imperceptible SynthID watermark, per the docs. When a client or platform asks where a track came from, the marker is in the file instead of in a spreadsheet someone forgot to update.

Still keep the prompt, the file name and the date next to each approved track. Check your ad platform's rules on disclosing AI-made audio too, since the watermark doesn't replace a label.

5. Safer-by-default prompts catch the "make it sound like..." request early

Prompts pass through safety filters, and the docs say requests for specific artists' voices or copyrighted lyrics are blocked. That catches the classic stakeholder note ("can it sound exactly like that famous singer?") at the API instead of in a legal review a week later.

The workaround is also the better brief: describe genre, tempo, instruments and mood. A blocked request is not a legal clearance for what does come back, so the human review below still applies.

When Lyria 3.5 is not worth it

Generation is single-turn. There is no multi-turn editing, so you cannot say "same song, shorter bridge." You rewrite the prompt and generate again.

Results also vary between calls with the same prompt. The take you loved cannot be reproduced by re-running it, so save every file you might want.

For brand-critical music, such as a sonic logo or a hero film, treat the output as a draft. Listen to the whole track, check sung words for odd pronunciation, and have a person sign it off. If the music is the whole point, hire a composer.

For social cuts, internal pitches and first drafts, eight cents a try is hard to argue with.

For a team shipping AI into a real workflow, cost, structure and provenance here are all things you can write into code and a review checklist, which makes the music something a team can approve instead of something one person tried once.