Google switched on Lyria 3.5, its newest DeepMind music model, inside the Gemini app and the Gemini API on 4 September 2026. Translation for marketers: type a sentence, get back a full track with vocals, verses and a chorus. No licence call. No composer invoice.
Google calls it its “best-sounding music generation model”, crediting more expressive vocals and richer arrangements. It is live globally on web and mobile, plus Google Flow Music, Google AI Studio and Google Vids.
What actually shipped
In the consumer Gemini app you get three new controls: pick or describe a genre, switch between vocal and instrumental, use templates, and set track length. The use cases Google itself names are commercial – a backing track for a video, a brand jingle, a personalised ringtone. Two of those three are marketing production line items.
The developer side is more precise. Lyria is a family, not one model:
- lyria-3-clip-preview returns exactly 30 seconds – built for loops, previews and prompt testing.
- lyria-3.5 generates songs of a couple of minutes with distinct verses, choruses and bridges.
- Both output 44.1 kHz stereo. MP3 by default; Lyria 3.5 can also return WAV.
- Up to ten images can accompany a text prompt, so a key visual can brief the audio.
- Prompts accept genre, named instruments, BPM, key, mood, structural tags like [Verse] and [Chorus], custom lyrics and timestamps.
That timestamp control matters more than it sounds. You can specify an intro from zero to ten seconds and a chorus from thirty to fifty – which is how a generated cue lands on your edit point rather than near it.
Limits worth knowing: generation is single-turn, so you cannot iteratively refine a clip across prompts. Identical prompts return different results. And safety filters block requests for specific artist voices or copyrighted lyrics.
Note the staged rollout
PPC Land flags a dating wrinkle worth repeating: Google announced Lyria 3.5 in Flow Music back on 29 July 2026, the API changelog listed it in public preview from 3 September, and the consumer post landed 4 September. Anyone treating a single date as “the launch” is describing half the story.
Why the distribution number is the story
Google Vids already had Lyria 3 and Lyria 3 Pro from 2 April 2026 – but behind an AI Pro or Ultra subscription, inside a workplace tool. The September change removes both the paywall and the product boundary.
Then look at reach. Alphabet reported 750 million monthly active Gemini users for Q4 2025, crossing 900 million by May 2026 and 950 million in Q2 2026. Dropping a music generator into that installed base does not change what is possible; it changes the volume.
For anyone selling stock music, this is a supply shock at the low end. A bespoke composer commission survives. The library cue bought for a fifteen-second social edit has a much harder pitch.
The compliance trap nobody sees
Every Lyria track carries a SynthID watermark – imperceptible, and Google says it does not affect listening. That satisfies the provider marking duty under Article 50 of the EU AI Act, which became legally applicable on 2 August 2026.
It does not satisfy the deployer duty. Deployers expressly include advertisers, agencies and publishers – and where labelling applies, it contemplates a visible disclosure, not an embedded signal. Penalties reach 15 million euros or 3% of worldwide annual turnover.
Practical consequence: a jingle generated in the consumer Gemini app and hand-uploaded to a paid social campaign inherits the marking but not the label. It also bypasses Google’s own “How this ad was made” disclosure panel, introduced 9 July 2026 across Google Ads, DV360, CM360, Merchant Center and Ads Editor, which auto-discloses assets made in Google tools.
The licensing backdrop
Generated music is not landing in calm waters. On 31 July 2026 the Munich I Regional Court ruled against Suno in the GEMA case, setting a 250,000 euro penalty per breach and rejecting fair use as a shield for training. In August 2026 Sony Music filed a 58.8 million dollar claim against Kroger over 392 songs used across the retailer’s social accounts.
That Kroger case is exactly why brand teams want Lyria. Unlicensed catalogue music creates a named counterparty with a sync-licence claim. A generated track has no such counterparty – but it carries unresolved questions about training data, and Lyria’s internal prompt rewriter and reasoning blocks are not exposed, so you cannot audit why a cue sounds familiar.
What to do this week
Test it on low-stakes assets first: organic social beds, internal decks, hackathon-style ad variants. Log which assets are generated. Add an AI-audio field to your asset management. And get legal to rule on visible labelling before a generated jingle runs in an EU paid campaign – because the interface will not ask you.
Source: PPC Land
