Lyria 3.5 Puts Full-Length AI Music in the Gemini App

Google Brings Its Music Model Into the Main App

Google announced on September 4, 2026 that Lyria 3.5, its music generation model, is now available in the Gemini app on web and mobile, in the Gemini API, in Google AI Studio, in Google Vids and in Flow Music. Until now Lyria has mostly lived in Flow, the tool aimed at artists and AI creatives. Putting it in the Gemini app is the change that matters, because it moves AI music generation from a specialist surface to the one most people already have open.

• Lyria 3.5 is live globally in the Gemini app on web and mobile, plus the Gemini API, AI Studio, Vids and Flow Music
• Full songs, not clips: Google's developer documentation lists the model as producing full-length songs with verses, choruses and bridges, running to a couple of minutes and steerable by prompt
• The older clip model stays as a separate option fixed at 30 seconds for loops and previews
• Every generated track carries a SynthID audio watermark, which Google says is inaudible and does not affect playback

Google describes Lyria 3.5 as its best-sounding music generation model, with more expressive vocals and richer arrangements than its predecessor. Those are the company's own comparative claims about its own model, so treat them as a vendor's characterization rather than an independent benchmark — there is no public head-to-head evaluation to point at yet.

What Changed for People Using the Gemini App

The practical upgrade is control and length. In the app you can pick or describe a genre, choose between vocal and instrumental styles, and select a short or longer track. Google has added templates aimed at common jobs — a backing track for a video, a jingle, a personalized ringtone — which is a sensible admission that most people asking for music do not know how to describe music.

The length change is the technically meaningful one. A 30-second clip is a loop. A track with verses, a chorus and a bridge has to hold structure across minutes, keep a key, and return to a theme without drifting. That is a harder generation problem, and it is the difference between a sound effect and a song. For developers, the API exposes both: a clip model fixed at 30 seconds, and Lyria 3.5 for full-length output in MP3 by default or WAV on request.

How Does Google Handle Provenance and Training Data?

Two things stand out on the responsibility side, and both are worth stating precisely.

First, watermarking. Google's documentation says all audio generated through the API carries a SynthID audio watermark, imperceptible to listeners, intended to let the track be identified as AI-generated later. That is the same provenance approach we looked at when Suno began watermarking its AI songs and when Apple Music announced Made With AI labels. Three different companies converging on audible-free, machine-readable provenance is a good sign for anyone who cares about knowing what they are listening to.

Second, training data. Google states that the Lyria model was trained only on licensed content. The company has not published a breakdown of which catalogs or which licensors, so this is an assertion by the vendor rather than a verified fact — but it is an assertion made on the record, and licensing has become the axis on which music-generation tools are being judged.

Who Is This Actually For?

The honest answer is people who need functional music and have no way to make it. A creator who needs 90 seconds of unobtrusive background for a tutorial, a small business that wants a jingle, a teacher assembling a slideshow — these are jobs that currently end in a stock-music subscription or in silence. Putting a competent generator inside Vids and the Gemini app aims squarely at that gap.

For musicians the calculus is different and more personal, and it is not one a news article should pretend to settle. What can be said neutrally is that the tool now sits alongside the search box rather than behind a specialist door, which means far more people will encounter it. For more on how Google's model family keeps expanding across surfaces, see our ongoing artificial intelligence coverage and our report on agentic video understanding in Gemini.

Sources: Google Blog — September 4, 2026; Google Gemini API documentation — accessed September 7, 2026; The Decoder — September 6, 2026.