Google has released Lyria 3.5, its AI music generation model, directly inside the Gemini app and through an application programming interface (API), according to The Decoder. The move puts Google’s music-generation tooling closer to everyday Gemini users while also making the model available to developers and creators through other Google products. The Decoder reports that Google says Lyria 3.5 improves on its predecessor with more expressive vocals and richer musical arrangements. In Gemini, users can specify genre and style, choose between vocal and instrumental outputs, and generate either short or longer tracks. The app also includes templates aimed at simpler use cases, including background music and personalized birthday songs. The rollout is not limited to Gemini. The Decoder says Lyria 3.5 is also available in Google Flow Music, where it has additional features, as well as in Google AI Studio for developers and Google Vids. That distribution suggests Google is positioning the model both as a consumer-facing creative feature and as a building block inside its broader AI tooling. The most sensitive claim is about training data. According to The Decoder, Google says Lyria was trained only on licensed content. The same report notes that Google has not shared specifics about what training data was used, leaving outside users and rights holders with the company’s high-level assurance rather than a detailed accounting. For builders, the API and AI Studio availability are the practical pieces: they make Lyria 3.5 usable beyond a single consumer interface. For Google, the Gemini placement gives the model immediate surface area in an app where users are already experimenting with text, image, and productivity tasks. Who benefits: Google benefits by adding another creative modality to Gemini and its developer stack. Gemini users, video creators, and developers get direct access to music generation without relying only on separate specialist tools. Who's exposed: Music-generation competitors are exposed where Google can bundle access into existing AI and productivity surfaces. Rights holders may still scrutinize the training-data claim because the provided report says Google has not disclosed the underlying dataset details.