Overview
Lyria 3.5 is Google DeepMind's music generation model. It debuted inside Google Flow Music on July 29, 2026, entered public preview through the Gemini API on September 3, and opened to all Gemini app users on September 4. It composes from a text prompt and outputs 44.1kHz stereo audio, with vocals, time-aligned lyrics and full instrumental arrangements.
DeepMind positions this generation as its most capable music model to date, with the improvements concentrated on vocal expressiveness and arrangement depth, alongside a clear jump in audio fidelity over Lyria 2.
Key Features
- Text and images as input: Prompts can describe genre, mood, instruments and structure. Up to 10 images can be supplied alongside a prompt, and the model will compose music inspired by what they show.
- Controllable lyrics and structure: You can write your own lyrics or have the model generate them around a theme. Section tags such as Verse, Chorus and Bridge shape the arrangement, and timestamps can specify what happens at a given moment, such as when the drums enter.
- From 30 seconds to 3 minutes: Two models are exposed through the API. `lyria-3-clip-preview` always produces 30-second clips for loops and previews, while `lyria-3.5` generates full songs with verses, choruses and bridges running a couple of minutes, with DeepMind's model page listing a three-minute ceiling.
- Output formats: MP3 by default, with a WAV option available for Lyria 3.5 so tracks can move into post-production.
- Language follows the prompt: Write the prompt in a given language and the lyrics come back in that language, with vocal style and pronunciation adapted to match.
- SynthID on every track: All generated audio carries a SynthID watermark, described by Google as imperceptible to the human ear, so platforms and listeners can tell whether music was created or edited with AI.
- Safety filtering: All prompts pass through safety filters, and requests for a specific artist's voice or for copyrighted lyrics are blocked. Before audio is generated, the model reasons through the musical structure implied by the prompt.
Use Cases
- Custom backing tracks for video, or a quick birthday song from a built-in template
- Brand jingles and personalised ringtones
- Original scores for short films by creators without a music budget
- Composing to a mood board by feeding in atmosphere images for a visual project
Pros
- 44.1kHz stereo with controllable structure is enough for real editing work
- Lyrics and section tags are steerable rather than left to chance
- Prompt language determines lyric language, so non-English creators can write natively
- SynthID watermarking on every output gives commercial use a paper trail
- Open to all Gemini app users, keeping experimentation cheap
Pricing
Available inside the Gemini app to all users on a standard account. Developers are billed through Google AI Studio and the Gemini API under Google's audio generation pricing, and enterprises can deploy through Vertex AI.
Summary
Lyria 3.5 moves music generation from demo-grade to something that can slot into a real editing pipeline. 44.1kHz stereo, controllable structure and language following make it serviceable for short-form video scoring and ad demos. Two current constraints matter: generation is single-turn, so iterating on one clip across multiple prompts is not the supported path, and identical prompts can still return different results between calls. Writing a detailed prompt and getting it right in one pass is the reliable approach.
Version History
- Gemini app general availability (2026-09-04): Opened to all Gemini users with genre selection, vocal or instrumental styles, short and long track options, and templates
- Gemini API public preview (2026-09-03): Developers can call `lyria-3.5` and `lyria-3-clip-preview`
- Lyria 3.5 debut (2026-07-29): Launched in Google Flow Music with improvements across musicality, lyrics and vocal quality