On September 4, Google put its Lyria 3.5 music generation model inside the Gemini app and the Gemini API. It looks like a version bump, but the real change is distribution: the model debuted in Google Flow Music on July 29 and spent five weeks as an exclusive capability inside a pro creative tool. Now it is a default option in a general-purpose chat box, plus a public endpoint any developer can call.

Five weeks, from pro tool to default option

As AI Musicpreneur recounted, Lyria 3.5 is not a new model — after launching with Flow Music on July 29, that was the only place you could use it for five full weeks. The September 4 move put it into the Gemini app, with Google's official wording that it is available "to all users globally" on both web and mobile. The same model went from "you have to seek it out" to "it is in the chat box" — a distribution lever most standalone music tools do not have.

New controls inside the Gemini app

The official announcement lists what you can do: select or describe a genre, choose between vocal or instrumental styles, use templates to jumpstart anything from background music to custom birthday tracks, and pick between short or longer tracks. Google calls Lyria 3.5 its "best-sounding" music generation model, with more expressive vocals and richer musical arrangements. Digital Trends adds that it builds full songs with actual verses, choruses and bridges rather than looping a short melodic clip — though once a track is generated, you cannot edit it mid-prompt.

The model keeps three other entry points: Flow Music for artists and AI creatives, Google AI Studio for developers, and Google Vids.

The API is the bigger half

Behind the Gemini API, any developer can call song generation from their own product. AI Musicpreneur's analysis names the difference from Suno and Udio: those are destinations — you go to the site, generate, leave. An API turns generation into plumbing that shows up inside apps that have nothing to do with music. Google's playbook is consistent as ever: one model, spread across every entry point of its own ecosystem.

Cold water and watermarks

The timing is not quiet. According to Digital Trends, AI-generated songs now make up 44% of daily uploads on Deezer, and a large share of the streams they pulled in was later found to be fraudulent. Every Lyria output carries a SynthID invisible watermark that flags it as AI-generated content — as music platforms get flooded, the watermark has gone from a compliance detail to infrastructure.

The templates point the same way: birthday tracks, ringtones and brand jingles are built for people who have never opened a DAW, not for producers.

The competition in AI music is changing axes: from "whose model sounds better" to "who shows up inside other people's apps". Lyria 3.5's model increment is modest, but its distribution increment is not — when music generation becomes a button in a chat box and one API call, standalone music tools need to re-measure their moats.

References: