You can now ask the Gemini app to whip up a birthday song. What you can’t get is the name of a single artist whose work taught it to do that.
That tension defines the Lyria 3.5 rollout: a music generation model shipping straight into the Gemini app and through the API, carrying a pointed claim about training data with nothing to back it up.
The licensing claim, and the hole in it
Per Google, the Lyria model was trained exclusively on licensed content — and the company draws that contrast directly against Suno’s music model. It’s a jab at the most obvious rival in generative music, the kind of line a firm only throws when it believes it’s standing on solid ground.
Yet no specifics about the actual training data have been shared. Not a catalog, not a label, not a rights holder. The assurance arrives without the evidence, so the claim is worth precisely as much as whatever trust you already extend to the company making it.
What you can do with it in Gemini
Inside the Gemini app, the flow is simple: choose a genre, choose a style, decide between vocals and instrumentals, and pick a short track or a longer one.
Templates ship alongside it, clearly built for people who have never prompted a music model — background music, or a personalized birthday song. Consider those starter rails rather than the ceiling of what the product can do.
What’s changed since the last model
Two upgrades anchor Google’s case for 3.5 over the previous model: more expressive vocals and richer musical arrangements. That framing is the company’s own, and it describes the sort of improvement nobody can really dispute and nobody can really confirm from a spec sheet either. There’s no number for expressiveness.
It’s not just the Gemini app
Lyria 3.5 also arrives in Google Flow Music, where Google says it brings added features. Developers reach it through Google AI Studio. And it shows up in Google Vids, dropping generated music into a workplace video tool instead of a creative sandbox.
That distribution is the real story, more than the model bump. A music generator living in a consumer chat app is a novelty. That same model threaded through a video editor and an API is Google angling to make generated audio the default filler for anyone who needs 30 seconds of something beneath a slide.
For anyone weighing whether to use it, the useful question isn’t whether the vocals have improved. It’s whether you’re willing to put this audio into published work when the company behind it insists the training data was licensed and won’t name one source.