Google Lyria 3.5 Review: Free Music Generation Inside the App You Already Use

Getting background music for a video used to mean the most annoying kind of errand. Create an account on a music generation service, pick a plan, register a card, and only then find out what the tool sounds like. By the time a single track plays, you have spent more effort on signup than on the creative decision.

On September 4, 2026, Google removed most of that errand. The company released Lyria 3.5, its music generation model, worldwide inside the Gemini app, AI Studio, and the API. A Japanese creator who writes about AI workflows tested it the same day, and his hands-on notes are worth reading closely, because his conclusion is not about sound quality at all.

What Lyria 3.5 actually is

Lyria is Google's music generation model family. Version 3.5, in the release the creator tested, supports composition from a specified genre, composition from an uploaded image, and vocal tracks at up to three minutes in 44.1kHz stereo.

On paper that reads as one more music model launch, and the numbers alone would make it forgettable. What he reacted to was the distribution. No dedicated service to register for. It sits inside the Gemini app, AI Studio, and the API, the places a working creator already has open. His phrase for the experience was close to "wait, it is already done," and that framing is the entire point.

The three tests

The genre test came first. He typed a plain request into the usual chat box, asking for a calm lo-fi hip-hop track for working. No mode to switch to, no separate screen. Generation started inside the conversation, and the response came back with an embedded player. One genre noun was enough to land the tempo and the overall feel, and he admits the smoothness exceeded what he expected.

The image test is the interesting one for anyone in visual work. He uploaded a photo of the sea at dusk and asked for a track that matched the photo, without describing colors, light, or texture in words. The model returned a quiet piano-and-strings piece that fit the image. His takeaway: passing intent through one photograph is more intuitive than decomposing a mood into adjectives. Anyone who has tried to verbally spec "warm but not cheesy, sparse but not empty" knows the pain he is describing.

The vocal test checked the ceiling. With lightly prepared lyrics, the model produced a full vocal track at the advertised three minutes and 44.1kHz stereo, with enough density to sit next to catalog music without apology. He is careful not to claim it equals a professional mix, but says the result no longer needs the excuse of being an AI demo.

Suno still wins when you are making music

The creator's comparison with Suno is the useful part of his notes, because it resists the "new tool kills old tool" reflex. Suno, which he had covered before around its v5.5 release, remains a dedicated music production service. Features like Voices, which lets you generate songs with your own voice, and Custom Models, which personalize v5.5 from uploaded catalog tracks, go deep into craft. If the goal is to produce music, Suno is still the right shelf.

Lyria 3.5, in his reading, is not really positioned as a music production app at all. Its natural job is the incidental track: background music for a video, a short cue under slides, a mood piece to attach to a photo. The value is finishing the task without leaving the app you were already in. For situations that do not justify registering anywhere, he expects Lyria to get the call far more often.

The caveats that survived the enthusiasm

Two notes from his testing keep the review honest.

Output character swings hard on word choice. Sometimes the first generation lands, and sometimes you regenerate several times with rephrased prompts before it does. Anyone who has worked with image generation will recognize the pattern: the model is sensitive to phrasing in ways that are hard to predict in advance.

And commercial use needs a terms check before client work. He flags licensing and copyright handling as something to verify against the actual terms before using a track in a paid engagement, rather than assuming free access means free rights.

Why distribution quietly beats features here

The deeper lesson from this launch applies well beyond music. For incidental creative assets, the track, the thumbnail, the quick visual, the dominant cost was never generation quality. It was the transaction cost around it: the new account, the new plan, the new card, the new export path.

Dropping that cost to zero inside the Gemini app changes behavior at the edges of a workflow, which is where most creator work actually happens. The one-off track that was previously not worth a signup now gets made, and the composite output, video plus music plus visuals, gets closer without any single tool needing to be the best at anything.

There is a second-order effect worth naming. When the incidental track is free inside the app, the floor for what counts as finished rises. A slideshow with silence now reads as unfinished next to one with a fitting cue, the same way a wall of plain text reads as unfinished next to one with visuals. Creators who ignore the audio layer are competing against outputs that silently got better.

How to approach it today

If you produce videos, slides, or photo pieces, spend ten minutes inside the Gemini app with Lyria 3.5 before your next incidental track errand. Test the image-to-music path first, since it is the least familiar and the most useful for visual creators: one photo, one sentence, one cue.

Keep Suno or a similar dedicated tool for anything where the music is the point rather than the accompaniment.

And before any client deliverable, read the current terms on commercial use. Free access and commercial rights are separate questions, and the answer lives in the license, not in the model.

The through-line of the Japanese review is hard to argue with. The achievement is not that another music model exists. It is that composition now happens where you already were, and for incidental music, that turns out to be most of the demand.