Quick Answer
Google has added Lyria 3.5 music generation to the Gemini app on the web and mobile devices worldwide. The feature creates vocal or instrumental tracks from a prompt, genre selection, or template. Google also makes Lyria 3.5 available through its API, AI Studio, Flow Music, and Vids. Review generated audio before sharing or using it commercially.
Key Takeaways
- Lyria 3.5 became available in the Gemini app and Gemini API on September 4, 2026.
- Gemini users can choose genres, vocal or instrumental output, templates, and track length.
- The API preview model generates 44.1 kHz stereo MP3 audio and text lyrics.
- Lyria 3.5 accepts text prompts or image inputs with a 131,072-token input limit.
- Google says the model includes filtering and SynthID watermarking safeguards.
What changed with Lyria 3.5 in Gemini?
Lyria 3.5 is now available in the Gemini app, giving consumer Gemini users a built-in way to generate music from prompts. Google announced the expansion on September 4, 2026, and said the feature is available globally in Gemini on the web and in the mobile app. Google’s announcement describes Lyria 3.5 as its best-sounding music generation model.
The Gemini rollout matters because Lyria 3.5 was initially introduced in Google Flow Music on July 29, 2026. The new availability moves the model from a more specialized music workflow into Gemini, where users can ask for generated audio alongside the other kinds of assistance the app provides. The expansion does not mean every Gemini feature is part of music generation, and users should treat Lyria output as generated media that needs review before publication.
Google’s Gemini product lineup continues to gain features across different surfaces, including the Gemini Flash rollout for Gemini, Search, and Sheets. Lyria 3.5 is distinct because its output is music rather than text, images, or spreadsheet assistance.
What can Gemini users create with Lyria 3.5?
Lyria 3.5 in Gemini can create short or longer tracks with either vocal or instrumental output. Google says users can select or describe a genre, use new templates, and choose the type of track they want. Those controls give a prompt more structure than a broad request such as asking Gemini to create a song without describing mood, instrumentation, or intended use.
Google positions Lyria 3.5 for 4 common use cases: custom backing tracks, brand jingles, birthday songs, and personalized ringtones. The feature is useful when a user needs an original-sounding audio concept for a personal project, but a generated track still requires a human check for tone, lyrics, and whether it fits the intended audience.
| Gemini music control | What the user chooses | Why the choice matters |
|---|---|---|
| Genre | A selected or described musical style | Genre gives the model a clearer direction for the track’s sound. |
| Output type | Vocal or instrumental | The choice determines whether the generated track includes sung content. |
| Template | A new music-generation template | Templates can provide a more defined starting point than an empty prompt. |
| Track length | A short or longer track | Length helps align the output with a ringtone, jingle, backing track, or song concept. |
The practical approach is to specify the genre, whether vocals are wanted, the desired length, and the purpose of the track. A request for an instrumental birthday ringtone needs different instructions from a request for a longer backing track, even when both start with the same musical style.
Where is Lyria 3.5 available beyond the Gemini app?
Lyria 3.5 is available through Google Flow Music, Google AI Studio, Google Vids, the Gemini app, and the Gemini API. Google says the Gemini app version is available globally on web and mobile, while the developer-facing version uses the preview model ID lyria-3.5. The different entry points matter because they serve different tasks, from consumer prompting to application development and video work.
Google Flow Music remains relevant because it was the first Google surface to receive Lyria 3.5. Google AI Studio and the Gemini API provide paths for users and developers who need to work more directly with the model, while Google Vids places music generation closer to video production. The available surfaces do not change the underlying need to inspect the result before adding it to a finished project.
Gemini users who use the service for personal organization should also distinguish music generation from features such as Gemini’s item-memory tools. Lyria 3.5 creates audio from supplied inputs, while item-memory features are designed to help users recall information such as where they placed an object.
How does the Lyria 3.5 Gemini API work?
The Lyria 3.5 Gemini API accepts text prompts or image inputs and returns 44.1 kHz stereo audio in MP3 format along with text lyrics. Google’s developer documentation lists lyria-3.5 as a preview model and sets a 131,072-token input limit. Google’s Lyria 3.5 developer documentation also identifies the output and supported capabilities.
The API specification matters because developers need to design around the model’s current limits rather than assume it supports every Gemini API feature. Google documents that Lyria 3.5 does not support live API, batch API, search grounding, function calling, or structured outputs. A developer building a music feature therefore needs to handle generated MP3 audio and lyrics without relying on those unsupported functions.
- Write a text prompt or provide an image input that establishes the desired musical direction.
- Send the request to the preview model ID
lyria-3.5. - Receive the generated MP3 audio and accompanying text lyrics.
- Review the audio, lyrics, and output quality before presenting the track to users or including it in a project.
Stop and consult Google’s developer documentation or support resources if an application requires live generation, batch processing, grounding, tool use, or structured output. Lyria 3.5 does not document support for those functions, so a workaround should not be assumed to be reliable.
What audio technology does Lyria 3.5 use?
Lyria 3.5 uses latent diffusion applied to temporal audio latents, according to Google’s July 2026 model card. In practical terms, the model generates audio through a system designed to work with compressed internal representations of sound across time. The technical approach helps explain why the model can produce audio from descriptive inputs, but it does not guarantee that every output will match a user’s prompt exactly.
Google says Lyria 3.5 was trained on captioned audio data. Captioned training data connects sound with descriptive information, which is relevant to a model that responds to prompts about genre, vocal output, or a particular mood. The model card does not make every generated track suitable for every context, and users remain responsible for reviewing whether an output meets their creative and practical requirements.
The most sensible approach is to provide clear descriptions rather than rely on a single vague label. A prompt that identifies the requested genre, output type, length, and intended context gives the model more direction, while still leaving room for generated variation.
What safety safeguards does Lyria 3.5 include?
Lyria 3.5 includes product-level safety mitigations, including filtering and SynthID watermarking, according to Google DeepMind’s model card. Google DeepMind’s Lyria 3.5 model card says the safeguards are part of the model’s product deployment. The protections matter because generated audio can be shared quickly and may be difficult for listeners to distinguish from conventional production without provenance information.
SynthID is Google’s watermarking technology for identifying AI-generated content. Watermarking and filtering can reduce some misuse risks, but they do not remove the need for responsible review by the person who creates or distributes the audio. Users should avoid presenting a generated vocal track as an authentic recording by a real person or using output in a way that confuses an audience about its origin.
Organizations should establish their own review process before using generated music in public-facing campaigns, customer communications, or branded materials. Stop and consult legal, policy, or communications professionals if a proposed use involves rights-sensitive material, impersonation concerns, or claims that could affect customers or the public.
How should users prompt Lyria 3.5 in Gemini?
Lyria 3.5 works best when a Gemini prompt gives the model specific creative constraints. Google confirms that users can select or describe a genre, choose vocal or instrumental output, use templates, and choose between short and longer tracks. The practical benefit of those options is that they help a user state the desired result before the model begins generation.
- Choose whether the track should be vocal or instrumental.
- Describe or select a genre that matches the intended use.
- Specify whether the track should be short or longer.
- Use a template when a predefined starting structure fits the project.
- Listen to the generated output and review any lyrics before sharing the track.
A useful prompt identifies the intended outcome instead of only naming a genre. For example, a user can request a short instrumental track for a ringtone or a longer vocal birthday song, then assess whether the finished result fits the request. Users should revise the prompt when the first output does not match the desired mood or structure.
Generated music also deserves the same careful handling as other AI-created media. Concerns about labels and transparency have become more common across platforms, including policies involving AI content identification on Instagram. Clear disclosure is especially important when an audience could mistake generated audio for a human performance.
Why does Lyria 3.5’s Gemini expansion matter?
Lyria 3.5’s Gemini expansion matters because it puts AI music generation in a consumer-facing app that is available globally on web and mobile. Google has also made the same model available through developer and production-oriented tools, including the Gemini API, AI Studio, Flow Music, and Vids. The broader availability gives individuals and developers more ways to create music, but it does not eliminate the judgment required to use generated output responsibly.
For most Gemini users, the immediate value is straightforward: a prompt can become a custom backing track, jingle, birthday song, or ringtone without separate music-generation software. For developers, the API adds 44.1 kHz stereo MP3 output and text lyrics to the available building blocks. The main limitation is that the preview API does not support several functions that developers may use elsewhere, including live API and structured outputs.
The practical response is to start with a clearly scoped creative task and evaluate the result before relying on it. Lyria 3.5 can make audio generation more accessible, while Google’s filtering and SynthID watermarking provide product-level safeguards rather than a replacement for human review.
FAQ
Is Lyria 3.5 available in the Gemini app?
Lyria 3.5 is available in the Gemini app on the web and mobile devices worldwide. Google announced the consumer rollout on September 4, 2026, alongside availability through the Gemini API.
Can Lyria 3.5 generate songs with vocals?
Lyria 3.5 can generate vocal or instrumental tracks in Gemini. Gemini users can choose the output type and can also select or describe a genre, use templates, and choose short or longer tracks.
What audio format does the Lyria 3.5 API return?
The Lyria 3.5 API returns 44.1 kHz stereo MP3 audio and text lyrics. The preview model accepts text prompts or image inputs under the model ID lyria-3.5.
Does Lyria 3.5 support live API or function calling?
Lyria 3.5 does not support live API or function calling under Google’s current developer documentation. The preview model also does not support batch API, search grounding, or structured outputs.
Does Lyria 3.5 include AI music safeguards?
Lyria 3.5 includes filtering and SynthID watermarking as product-level safety mitigations, according to Google DeepMind. Users should still review generated tracks before sharing them, especially when the output could be mistaken for a human recording.