What shipped
Google has made Lyria 3.5, its music generation model, available in the Gemini app and through the Gemini API. The company announced the move on 4 September, describing more expressive vocals and richer arrangements than earlier versions.
The model produces 44.1kHz stereo audio and full song structures — verses, choruses and bridges rather than a looped phrase — at lengths up to three minutes, according to Unite.AI’s account of the launch. Users can write their own lyrics or have the model generate them around a subject, and can specify vocal style and genre. A separate variant, Lyria 3 Clip, produces fixed 30-second clips.
Access is broad. Google says the model is available to all Gemini users globally on web and mobile, and it also sits in Flow Music, Google AI Studio and Google Vids. Google has not published API pricing.
Why the distribution matters more than the model
Lyria 3.5 is not new. It launched in Flow Music on 29 July 2026, aimed at people who had gone looking for a music tool. What changed on 4 September is where it lives.
Music generation has until now been a destination product: you went to Suno, or Udio, or Flow, because you wanted to make a track. Putting a competent three-minute-song model inside a general assistant that hundreds of millions of people already open for other reasons removes the step where someone decides they are the sort of person who makes AI music.
That is the same manoeuvre Google has run with image generation, and it is the one that most reliably moves usage without moving quality.
The watermark, and what it does not settle
Google says every track carries an imperceptible SynthID watermark, so generated or edited audio can be detected afterwards.
That matters more than it did a month ago. The EU AI Act’s transparency obligations, enforceable since 2 August 2026, require machine-readable marking of AI-generated content, and California’s AI Transparency Act has been operative since the same date, obliging large providers to offer detection tools and provenance disclosures. A model shipping into a globally distributed consumer app is squarely in scope of both.
Watermarking answers the labelling question. It does not answer the training one. Music publishers are currently litigating what a model may be trained on — Sony Music Publishing and Warner Chappell sued Anthropic over song lyrics in late August — and Google has not said what Lyria 3.5 was trained on.
What to watch is API pricing, which will determine whether Lyria 3.5 competes with Suno and Udio on the developer side or only inside Google’s own surfaces.
