Google puts Lyria 3.5 into the Gemini API for full-song generation

Google is bringing Lyria 3.5 into the Gemini API public preview, extending its latest music-generation system beyond the Flow Music product surface. The model is positioned for full-length songs with more coherent musical structure, improved vocals and lyrics, text or image input, and controllable duration.

The release matters to developers because music generation is becoming an API capability rather than a closed creative demo. It also comes with a practical warning: Google's live documentation has not exposed one perfectly consistent model naming surface, so teams should confirm the identifier and pricing page they will actually call.

Lyria 3.5 moves from Flow Music to an API

Google DeepMind describes Lyria 3.5 as its newest music model, with improvements in musicality, lyrics, vocal quality, pronunciation, and creative control. Its product page shows tracks of up to three minutes and image-guided composition, allowing a visual reference to shape the generated music.

AI Stack Current reports that the Gemini API public preview began on September 3, 2026. The report identifies full-song generation, vocal and lyric output, prompt-controlled structure, and paid per-request access as the main developer-facing changes. That makes this a new access stage for an existing model family, rather than a claim that the underlying research model first appeared this week.

The model ID needs a defensive check

The current documentation uses lyria-3.5 for the full-song route in some surfaces, while other live pages have shown launch-era names such as lyria-3.5-pro-preview and lyria-3.5-clip-preview. AI Stack Current flags the mismatch across Google's release notes, model references, and music-generation guides.

Before changing an application, developers should verify the model ID returned by the current model list, check whether the request uses the GenerateContent or Interactions API, and confirm the output format and per-request price. A copied identifier from an older guide can fail even when the model itself is available.

What creators and builders can use now

Lyria 3.5 is aimed at structured music creation: verses, choruses, bridges, vocal direction, genre and instrumentation prompts, and variable song length. Image input broadens the workflow for trailers, mood boards, social clips, and game or video concepts. The model page presents these as capabilities; it does not establish independent quality rankings against Suno, Udio, or other generators.

For production work, treat the preview label as a real constraint. Preview model behavior, limits, endpoint names, and output policies can change. Keep the model identifier configurable, log the exact API response metadata, and retain a fallback route for short clips if a full-song request is unavailable or too expensive.

Evidence and the next integration step

The official Google DeepMind Lyria page establishes the model's current capabilities and product positioning. The independent AI Stack Current report corroborates the September Gemini API preview and documents the endpoint-name conflict. Together they support the release stage while leaving implementation details to be checked against the live Google developer console.

If you are building a broader agent or creator workflow, LinkLoot's AI workflow automation guide is a useful place to map generation, review, storage, and publishing steps before adding a preview API to a production pipeline.