What does Lyria 3.5 change for anyone making music in Gemini?

The model inside the chat window

Lyria 3.5 is a music generation model. Google calls it their best-sounding one yet. It now lives in two places: the Gemini app that anyone can open and the Gemini API that developers call from code.

The announcement says the model brings more expressive vocals and richer musical arrangements. Higher fidelity is the result. In practice this means the audio files sound less like a demo and more like a finished track. The model does not just hum a melody. It structures verses and choruses. It also decides where the drums kick in and chooses chord voicings that fit the genre you name.

Controls that shape the output

The Gemini app adds a row of controls above the prompt box. You can select a genre from a list or describe one in your own words. You can choose between vocal and instrumental styles. The switch is binary. Vocals on. Vocals off. No slider for amount of singing.

Templates appear next to the genre picker. They cover background music and custom birthday tracks. Each template pre-loads a prompt structure so you do not start from a blank page. Track length is now a choice. Short or longer. The announcement does not give seconds or minutes for either setting. It only says you have the flexibility to choose.

A custom backing track for a video is one use case. A unique brand jingle is another. A personalized ringtone is a third. The app renders the file and hands it to you inside the chat thread. You can play it. You can download it. You can prompt again if the bridge feels weak.

Three doors to the same engine

The same Lyria 3.5 model powers three different products. The Gemini app is the consumer door. It is free and global. Google Flow Music is the artist door. It targets artists and AI creatives who want deeper control over stems and arrangement data. Google AI Studio and Google Vids form the developer door. AI Studio lets technologists prototype prompts and chain the model with other tools. Google Vids embeds the engine inside a video creation workflow so a soundtrack appears while you edit clips. Each door shows a different slice of the API. The underlying weights do not change.

What the announcement leaves open

No word on the training data. No word on copyright indemnity for commercial releases. No audio format specifications such as sample rate or bit depth. No maximum duration in seconds for the longer setting. No pricing tiers for the API beyond the implication that AI Studio access exists. No latency figures for real-time streaming. No offline or on-device option. The document simply states the model is available now. It describes what you can do today. It does not describe what you cannot.