Google Opens Lyria 3.5 to Developers and Adds Song Generation to Gemini

Google has put Lyria 3.5, its latest music-generation model, into the Gemini app and Gemini API, giving both everyday users and developers a route to generate vocal or instrumental tracks. The consequential change is distribution: a tool previously limited to Flow Music is now available through Google AI Studio, the API, and Gemini’s consumer interface.

From a contained music tool to wider access

Google’s public announcement describes Lyria 3.5 as a higher-fidelity update with more expressive vocals and richer arrangements. In Gemini, people can select or describe a genre, choose vocal or instrumental output, use templates for background music or birthday tracks, and choose shorter or longer tracks. The company says the feature is available globally on the web and in Gemini’s mobile app.

For builders, the expansion creates a programmatic path alongside the app. Google says Lyria 3.5 is also available in Flow Music and Google Vids. Google’s API documentation lists the two Lyria 3.5 preview model IDs with a September 3 release date, and does not list a shutdown date for either one.

Two generation paths

01Fixed 30-second MP3

Lyria 3.5 Clip

Clip is positioned for fixed 30-second outputs.

02Multi-minute songs

Lyria 3.5 Pro

Pro generates multi-minute songs with verses, choruses, and bridges; prompts control duration.

Controls reach beyond a single audio file

The two variants can generate 44.1 kHz stereo audio from text prompts or from a text prompt paired with as many as 10 input images. The developer-facing controls include custom lyrics, multilingual vocals, section tags, timestamps, and prompt-level direction over tempo, duration, and key.

What the API returns

• Audio blocks containing the generated music.
• Generated lyrics alongside a JSON breakdown of the song’s structure.

That structured response can make the model more useful for an application that needs to display both a track and its lyrics or sections, rather than treating the result as audio alone. But the available controls shape the initial generation, not an editing workflow after the fact.

The production boundary is still regeneration

Lyria 3.5 is single-turn: changing a chorus or another part means regenerating the whole track. It does not expose seed control, so developers cannot request the same generated result again through a fixed seed. Those constraints distinguish prompt specification from fine-grained revision, a material issue for teams trying to make a particular draft reproducible.

Google also applies guardrails at the request and output stages. Every Lyria 3.5 track carries its SynthID audio watermark, while requests for artist mimicry and copyrighted lyrics are blocked. Google says Lyria was trained on licensed data from material that YouTube and Google had rights to use; that is the company’s account of its training-data basis, not an independently detailed dataset disclosure.

The next question is how quickly the preview becomes a stable developer product and whether Google adds editing or reproducibility controls. For now, the company has widened access to full-song generation, but it has not set a retirement date for either preview endpoint.