This is more than a new prompt box
Google has released another music model. The more interesting part is where it put it.
Lyria 3.5 began rolling out on Wednesday inside Google Flow Music, a workspace built around making and revising songs. A user can start with text, an audio reference, a recording or an image. The result can then be edited, remixed, extended, covered or partly replaced.
That makes the product feel less like a one-shot generator and more like a small studio. Songs can be organised, shared and downloaded. Flow Music also includes playlists, collaborative spaces and tools for making music videos.
The service is for adults in supported regions. Google offers a free sign-up and sells additional credits through Flow Music or Google AI subscriptions. Its public help pages do not present one simple global price because access and plans vary.
The shift matters. The model still generates the sound, but the product is now built around what happens after the first attempt.
What Google says Lyria 3.5 improves
The launch note is short and specific about the areas Google wants people to notice.
The company says Lyria 3.5 makes richer and more natural melodic structures. It also claims better lyrics, stronger adherence to the prompt and more awareness of how a song is organised.
Vocals are meant to sound more expressive, with better pronunciation. Users also get more direct control over tempo and the duration of an output.
These are company claims. Google has not published a listening test, a side-by-side benchmark or a detailed comparison with Lyria 3. There is no public sample showing how often the requested tempo, structure or lyric actually survives a full edit.
Music quality is also hard to reduce to one score. A cleaner voice can be easy to hear. Whether a song feels less generic is a different and more subjective question.
Editing becomes the real product
Flow Music's controls make the direction clearer than the launch language does.
A creator can ask for a shorter introduction, a different mix, another set of instruments or a change of style. A selected section can be extended or replaced without throwing away the whole track.
That is useful because generated music rarely arrives finished. A good chorus can sit next to a weak verse. The voice can work while the arrangement drifts. Keeping the useful part is closer to how people actually make things.
Google also lets users publish a song by link and decide whether other people may remix or download it. Those are ordinary product choices, but they turn generation into a small distribution system too.
The open question is how much control the tools really provide once the novelty wears off. A list of editing commands is not the same as predictable editing across genres, languages and long sessions.
Every track carries a Google watermark
Google says all music made with Lyria is marked with SynthID, its inaudible watermark for generated audio.
The watermark is embedded in the sound rather than attached only as metadata. Google says it is designed to remain detectable after common changes such as MP3 compression, added noise or a change in playback speed.
People can upload an audio file to Gemini and ask whether Google AI made or edited it. Google also has a separate SynthID Detector in testing with journalists and media professionals.
That helps identify output from Google's own system. It does not prove that an unmarked track was made by a person, and Google itself describes SynthID as one building block rather than a complete answer.
The model card says Lyria uses filtering, safety reviews and product safeguards. It does not list the music used for training. The Lyria 3 launch said the system was designed for original expression rather than mimicking named artists, but the new 3.5 note adds no fresh evidence on that question.
What is confirmed, claimed and still open
Confirmed: Lyria 3.5 is rolling out in Flow Music. The product supports text, image and audio references, editing, remixing, sharing and downloads. Google says Lyria output receives a SynthID watermark.
Google's claims: the new model improves musicality, lyrics, prompt adherence, vocals, pronunciation and control over tempo and duration. No public evaluation accompanies those claims.
Still open: the training-data mix, failure rates across languages and styles, the practical limits on commercial use in each plan and how reliably edits preserve the parts of a song a creator wants to keep.
None of that makes the release trivial. Putting generation, revision and distribution in one place is a more meaningful product move than adding another quality adjective to a model page.
The fair test is simple and slightly unglamorous: can someone return to a rough song, change one thing and keep the rest? If Flow Music can do that consistently, Lyria 3.5 will feel less like a demo and more like an instrument.
Sources
- Google — Introducing Lyria 3.5 in Google Flow MusicPrimary launch note published 29 July 2026. Source for availability and Google's claims about musicality, lyrics, vocals, pronunciation, tempo and duration controls.
- Google Flow Help — Create and manage songsPrimary product documentation used to verify text, image and audio inputs, editing, remixing, sharing and download controls.
- Google Flow Help — Get started with Flow MusicPrimary availability and account documentation used to verify the age requirement, supported-region condition, free sign-up and paid credit options.
- Google DeepMind — Lyria 3 model and model cardPrimary technical and safety documentation. Google says the card covers Lyria 3 and subsequent versions and describes evaluations, filtering and SynthID use.
- Google DeepMind — SynthIDPrimary documentation for audio watermarking, stated resistance to common audio modifications and available detection routes.
