Google has added AI music generation to the Gemini app, running on a new model called Lyria 3.5. The same model is available in Google Flow Music, where it carries additional features, in Google AI Studio for developers, and in Google Vids. Google says Lyria 3.5 was trained only on licensed content, and drew the contrast with Suno's music model by name. It did not disclose what that licensed content actually is.
Introducing Lyria 3.5 music generation in Gemini
Source: the-decoder.com
The distribution is the news here, more than the model. Four products at once — the assistant, Flow Music, the developer platform, Vids — is not how a company ships an experiment. It is how a company ships a component it expects other things to be built on. The version number points the same way: 3.5 is an increment, not a debut, and the change worth noticing is not the model's capability but the fact that Google decided it was ready to sit underneath everything.
The tiering is quietly stated and worth reading. The Gemini app gets music generation; Flow Music gets music generation with additional functions. In other words, the version most people will touch is the sampler, and the controls live in the tool aimed at people actually making something. My read is that this split is deliberate rather than a staging accident: the broad surface builds the habit, the specialized surface holds what a serious user would pay attention to.
Then there is the licensing claim, which is the part of the announcement doing the most work. Google says the training data was licensed, and it says so by naming Suno. That is not a product argument. A user picking between two music generators does not care about a training corpus; a legal department deciding whether its company can put generated audio into a shipped video does. The sentence is aimed at procurement, and it is the reason Lyria 3.5 is in Vids and AI Studio and not just in the chat app.
Which is why the second half of the announcement undercuts the first. Google asserts a clean corpus and declines to say what is in it. A licensing claim with no disclosure is a claim you are asked to take on trust, and trust is precisely what the claim is supposed to replace. Naming a competitor while withholding your own evidence reads like a company that wants the reputational benefit of being the licensed option without accepting the auditability that would make the label mean something. If the corpus is as clean as Google says, describing it costs nothing.
The question nobody asked in the announcement is what "licensed" covers. Licensed from whom — publishers, labels, production-music libraries, individual artists? Licensed once, or with ongoing payment tied to output? Licensed for training only, or for commercial use of what the model produces? Those are four different legal situations that share one adjective. A musician cannot tell from this whether their work is in the model. A company shipping Lyria 3.5 audio in a product cannot tell what, exactly, it is indemnified against.
That gap matters more as the model spreads. Audio generated in a chat app is a toy; audio generated in Vids and through AI Studio ends up inside other people's finished work, distributed under their names. Google has attached a promise to that output and kept the paperwork behind it private. The first time someone has to show which license covers a specific generation, "trained only on licensed content" stops being a line in a blog post and becomes something a court or a customer expects to see documented.