What Is Lyria 3.5? Google's Music Model, Now in the Gemini API

Lyria 3.5 is now a step a pipeline can call

Lyria 3.5 is Google's music generation model. It writes songs from a text prompt and returns the audio plus the lyrics. It sings. It does not narrate. Google has shipped it in three dated steps in just over five weeks: Google Flow Music on July 29, 2026, the Gemini API as generally available on September 3, and the Gemini app on September 4. In the API the model code is lyria-3.5, and Google's pricing page lists it at $0.08 per song on the paid tier, as read on September 21, 2026.

That API line is the one that matters if you make ads.

Picture the end of a normal edit. Twenty-two seconds of a person holding a serum bottle, the voiceover timed to the product reveal, captions burned in, and a silent gap where the music bed goes. When Lyria 3.5 launched, filling that gap with it meant opening Flow Music, auditioning takes, exporting a file and dragging it into the timeline by hand. Since September 3 it can be one more request in the same pipeline that renders everything else.

Two things before the details. This is a music model, so it does not replace the voiceover layer of an ad. And Novoads does not run Lyria. What follows is a dated read of where the model lives, what the API offers, what it costs, and what that changes for an ad soundtrack.

Where you can use Lyria 3.5 today

The dated timeline

Google opened Lyria 3.5 in stages, and each step has its own dated Google page:

• July 29, 2026: Google Flow Music. Google's launch post called Lyria 3.5 its "newest music generation model" and rolled it out that day in Flow Music.
• September 3, 2026: the Gemini API. The Gemini API release notes list "Lyria 3.5 generally available (GA)" under that date, with the model code lyria-3.5.
• September 4, 2026: the Gemini app. A Google post that day says Lyria 3.5 "is now available in the Gemini app and the Gemini API", and that it is available to all users globally on the web and in the mobile app.
• Google AI Studio and Google Vids. The same September 4 post names both as places developers and technologists can use the model, without a separate launch date.

"Available to all users globally" is Google's own claim. This post did not check regional exceptions.

Which surface fits which job

The surfaces are not interchangeable. Each suits a different kind of work:

• Flow Music is a studio app, aimed in Google's words at artists and AI creatives. You generate, audition and export by hand.
• The Gemini app is the quick route. Google says you can select or describe a genre, choose between vocal or instrumental styles, start from templates, and choose short or longer tracks.
• The Gemini API is the surface a pipeline can call, the same one that serves Google's Veo 3.1 Lite video model. It is the one with a model code, a stable version and a published price per song.
• Google AI Studio sits in between: a place to try prompts against the API model before you write any code.

Why the model card still names one channel

If you read the DeepMind model card, the picture looks narrower. The card is marked "Published 29 July 2026" and "Updated: July 2026", and its Distribution section names one channel: Google Flow Music. It also lists text as the only input, while the API documentation lists text and image.

My reading: the card predates the API launch and has not caught up. For the question of where you can use the model, the newer dated pages win, which are the September 3 release notes and the September 4 announcement.

Lyria 3.5 in the Gemini API: model code, inputs and price

Model code and status

The model code is lyria-3.5. The model page lists a single version, "Stable: lyria-3.5", and the September 3 release note calls it "generally available (GA)". The same note describes Lyria 3.5 as "the next generation of Google's music generation model".

That is a release, not a preview. If you are deciding whether to build on it, that label is the first thing to check.

What goes in and what comes out

The model page and the release notes together make a short spec sheet:

• Inputs: text and image.
• Output: "Audio (MP3), Text (Lyrics)", so you get the track and the words it sings.
• Audio: 44.1 kHz stereo, per the release notes.
• Shape: full-length songs "including multiple verses, choruses, and bridges", per the model page.
• Control: the release notes promise "fine-grained duration and structural control".
• Batch: the model page marks the Batch API as "Not supported".

That last line deserves a second look if you plan volume. My reading: a round of twelve variants is twelve ordinary requests, not one batch job, so plan retries and rate limits around single calls.

What it costs per song

Google's Gemini API pricing page lists Lyria 3.5 (Full Song) at $0.08 per song on the paid tier. Its free-tier column for the model reads "Not available". The page was last updated on September 16, 2026, and this price was read on September 21, 2026. Pricing pages move, so check it before you budget.

The variant math: twelve variants at one song each is 12 × $0.08 = $0.96. Ask for three takes per variant so you have something to choose from, and it is 36 requests, or $2.88. At list price, the music for a full round of creative testing now costs under three dollars. The expensive part is the listening.

What Google says changed in Lyria 3.5

The four advances, in Google's own words

Google's July 29 post frames Lyria 3.5 as delivering "significant advancements across musicality, lyrics, and vocal quality". Underneath sit four specific lines:

• Musicality. Richer, more complex melodic structures that sound more natural. The advertiser's read: fewer beds that audibly loop at the fifteen-second mark.
• Lyrics. Higher quality lyrics with improved prompt adherence and structural awareness. The read: a sung hook you can steer toward the product instead of away from it.
• Vocals. More realistic and emotionally nuanced vocals, plus improved pronunciation. The read: a vocal in your market's language that does not sound machine-read.
• Creative control. More easily control the tempo and duration of your outputs. The launch post adds no numbers to that line.

The API release notes describe the same idea as "fine-grained duration and structural control", and DeepMind's Lyria page advertises tracks "up to 3 minutes long".

Two superlatives run through Google's copy: "our best-sounding music generation model" in the September 4 post, and "most capable music generation model to date" in the model card. Both are Google describing its own product, not a test result.

What the model card measures, and what is marketing

The model card is written for a different reader, and it is more useful if you care about what shipped rather than how it feels. On Google's own evaluation, Lyria 3.5 "improved significantly compared to Lyria 2 on audio fidelity", and with lyrics it shows better prompt adherence on both simple and more complex instructions.

Sort the claims by the document that carries them and a pattern shows up:

• Evaluation sentences: the audio fidelity gain over Lyria 2 (model card) and better lyric prompt adherence (model card and launch post).
• Marketing sentences: richer melodic structures, improved pronunciation, and easier tempo and duration control (launch post).
• Mitigation: SynthID watermarking (model card and DeepMind's Lyria page).

Musicality and vocals get marketing sentences. Lyrics and fidelity get evaluation sentences. And the detail a working advertiser should care about most, the watermark, appears in neither announcement post.

Training data, and the limit Google admits

On training data, the card says Lyria 3.5 "was trained on audio data" and that the audio datasets were annotated with text captions at different levels of detail. That is the whole disclosure. If you have followed the litigation around AI music training sets, you already know why a sentence at that altitude is not a chain of title.

DeepMind also publishes an honest caveat on its Lyria page: the team is "still working on improving key capabilities", and you should check that the tracks you create fit your vision. That is a fair thing to say about any generative audio in 2026.

Lyria 3.5 vs Lyria 3

Two previews, now labelled legacy

Lyria 3.5 is not Google's first music model in the API. On March 25, 2026, the release notes record the launch of two Lyria 3 models: lyria-3-clip-preview for 30-second clips and lyria-3-pro-preview for full-length songs. Both take text and image inputs and generate 48kHz stereo audio, per that entry. Google's pricing page now describes Lyria 3 as "Google's family of legacy music generation models."

Side by side

Model	Status on Google's pages	Output	Paid price
lyria-3-clip-preview	Preview, labelled legacy	30-second clips	$0.04 per song
lyria-3-pro-preview	Preview, labelled legacy	Full-length songs	$0.08 per song
lyria-3.5	Stable, GA since Sep 3	Full-length songs	$0.08 per song

Prices are from Google's pricing page as read on September 21, 2026. All three list the free tier as "Not available".

Which one to call for an ad

My pick for new work is 3.5. It costs the same per song as the Lyria 3 pro preview, it is the version Google marks stable, and it is the one Google put into the Gemini app and Flow Music.

The only case for the old clip model is price: half the cost per song, and a fifteen-second cut does not need a full song. But a legacy preview is a dependency on a model its maker has already moved past, which is a poor thing to wire into a pipeline you want running next quarter.

What Lyria 3.5 changes for an ad soundtrack

Google now names the ad jobs itself

Google's September 4 post lists the uses it has in mind, among them "a custom backing track for a video, a unique brand jingle". The Gemini app also lets you choose between vocal and instrumental styles.

Read together, those two lines cover most of what an ad asks of music: an instrumental bed under a spoken read, or a short sung hook for a brand. DeepMind adds that Lyria can generate vocals in different languages, which starts to matter the moment the same ad runs in a second market.

The bed was the last manual step

Look at what an AI ad pipeline already does with nobody touching it:

• Script. Drafted from a product page or a one-line brief, in the language you chose.
• Actor. A person who does not exist, delivering that script to camera.
• Product imagery. Generated from one uploaded photo, at whatever angle the shot needs.
• Voice. Rendered straight from the script text, with open-source voice models now in the mix.
• Assembly. Cuts, captions, and a reframe for each platform's spec sheet.

Then it used to stop and hand you a nearly silent file. Video models had started to generate synchronized audio in the same call as the picture, but music sat in its own app, behind its own front door. Teams that run structured creative operations know what a manual step inside an automated cadence does: it sets the pace for everything around it.

A stable music model with a model code and a price per song removes that excuse. The bed is now a choice, not a chore.

Music becomes a variable you can test

A stock bed used to be a constant by necessity, because nobody sources twelve different tracks by hand for one test. A callable model turns music into something you can test on purpose. A workable order:

1. Hold one generated bed constant while you test hooks, so the music cannot explain the winner.
2. Take the winning cut and generate three beds for it, instrumental and vocal, at the cut's exact length.
3. Test the beds against each other and against no bed at all. Silence is a valid arm of the test.

At list price, step 2 costs $0.24 per winning cut.

Length that matches the cut

The oldest indignity of ad audio is trimming a full-length song down to twenty-two seconds and hoping the fade lands somewhere musical. DeepMind's Lyria page promises to "generate cohesive tracks matching your project's exact duration", and the API release notes promise duration control. If that holds in your own tests, you ask for the length of your cut and get a piece that resolves there, instead of an excerpt that just stops. For a format built on tight cuts, that is a bigger change than better melodies.

A watermark you can point at

The model card lists SynthID watermarking among its product-level mitigations, and DeepMind's Lyria page states that all of its tracks are "imperceptibly watermarked with SynthID technology", so music created or edited with AI can be detected.

Read that as an advertiser rather than as a musician. The hard question about a generated track has never been "will anyone notice". It is "what will you say when someone asks where this came from". A documented watermark is a partial answer, and a partial answer beats a shrug. It cuts both ways: detectable by you means detectable by everyone.

When the best bed is no bed

Here is the position I will not hedge: most creator-style ads sound better with no music at all.

A bed is a polish signal, and polish is exactly what the UGC-style format is built to avoid. A paid video that sounds produced stops reading as a recommendation and starts reading as an ad.

Keep the bed when:

• No voiceover is carrying the timing, so the audio has to do the pacing.
• The piece is a montage, a product-motion cut, or a fashion film with no dialogue.
• The ad will run for months, so a careful choice pays for itself.

Cut the bed when:

• A person is talking to camera for most of the runtime.
• The ad is meant to land as a recommendation rather than as a commercial.
• You have not yet tested the version with no music.

Either way, before a generated track runs in paid media, read the terms of the Google surface you generated it on; this post does not summarize them. If you would rather not generate at all, a licensed production-music track with the license file saved beside the creative, or your ad platform's own cleared sound library, still works.

How Novoads handles the layers that carry the message

Novoads generates the layers of a UGC-style ad that carry the message, from a script you write or auto-generate:

• The video, from an uploaded product image or a prompt.
• The AI actor, delivering the script to camera. To show your product in hand, you build a custom actor from a photo of someone holding it.
• The voiceover, in voices spanning 31 languages.

Novoads' model catalog also includes a Suno music model. Novoads does not run Lyria, so if you want a Lyria 3.5 bed, generate it in the Gemini API, the Gemini app or Flow Music, add it in the edit, and keep the terms you generated under beside the asset.

On cost, a finished clip in Novoads runs from about 25 cents for a five-second Seedance 2.0 Mini clip to about $4 for an eight-second Google Veo 3.1 clip, and about $10 for a full 30-second Seedance 2.5 take. Every plan is published on the pricing page.

The bottleneck moved from access to taste

In July, Lyria 3.5 launched inside one app. By September it was a stable API model with a model code and a published price per song. For ad teams, that move is the real news, more than any line about melodies.

It also removes the last good excuse for mediocre ad audio. Soundtracks were an afterthought not because the available music was bad, but because sourcing it was manual and nobody could say where it came from six months later. A callable model with a documented watermark makes both easier to solve.

What is left is judgment: whether this ad needs music at all, and which of three beds fits the cut. A music model you can call makes the bed cheap. It does not make it necessary.