Google made Lyria 3.5, the newest model in its DeepMind music generation family, available inside the Gemini app and the Gemini API on September 4, 2026, extending machine-composed audio from a specialist creative tool to the company's largest consumer artificial intelligence surface and to any developer holding an API key.

In short

Google has switched on its newest music model inside the Gemini chatbot and opened the same model to outside developers through its programming interface. That means anyone with a Google account can now describe a song in words and get back a finished track with singing, and any software company can wire that capability into its own product. Every track generated this way carries an inaudible marker identifying it as machine-made, which is the detail that matters most for anyone publishing the output commercially in Europe.

What Google shipped

The announcement, published on the Google blog under the Gemini App category and dated September 4, 2026, describes Lyria 3.5 as available in both the Gemini app and the Gemini API. Google characterises it as its "best-sounding music generation model" and attributes the improvement to more expressive vocals and richer arrangements.

Distribution is unusually broad for a first release. According to the announcement, Lyria 3.5 reaches all users globally, on the web and in the mobile application. The same model is available to artists and creatives through Google Flow Music, and to developers through Google AI Studio and Google Vids. That is five distinct product surfaces named in a post running to fewer than two hundred words.

There is a dating wrinkle worth recording rather than smoothing over. Google published a separate post on July 29, 2026 announcing that Lyria 3.5 was rolling out in Google Flow Music, with improvements described across musicality, lyrics, vocals and creative control. The Gemini API changelog lists the model in public preview from September 3, 2026. The September 4 consumer post presents the model as newly available without referencing the earlier Flow Music release. Coverage that treats either date as the launch describes half of a staged rollout.

Inside the Gemini app

Three changes land in the consumer interface. Users can select or describe a genre and choose between vocal and instrumental styles. A template system arrives, which Google frames as covering everything from background music to custom birthday tracks. And track length becomes adjustable, with the post noting the flexibility to choose short or longer outputs.

The use cases Google names are commercial ones. The post cites a custom backing track for a video, a brand jingle and a personalised ringtone. Two of those three sit squarely inside marketing production. A jingle is a piece of sonic branding; a backing track is the audio bed under a social video or a pre-roll unit. Neither has historically been free, and neither has historically been produced without a licensing conversation.

Google Vids already carried a version of this capability. The April 2, 2026 expansion of Vids brought Lyria 3 and Lyria 3 Pro into the editor for Google AI Pro and Ultra subscribers, producing tracks between 30 seconds and three minutes without leaving the editing environment. That was a paid tier inside a workplace product. The September 4 change removes both the subscription gate and the product boundary.

The developer surface

The Gemini API documentation, last updated September 4, 2026, sets out what the model does at a technical level, and the detail is more specific than the consumer post.

Lyria is documented as a family rather than a single model. Lyria 3 Clip, exposed as lyria-3-clip-preview, always returns exactly 30 seconds and is positioned for loops, previews and prompt testing. Lyria 3.5, exposed as lyria-3.5, generates songs lasting a couple of minutes with distinct verses, choruses and bridges, and duration can be steered through the prompt or through explicit timestamps written into it. Both run through the Interactions API and both produce 44.1 kHz stereo audio. MP3 is the default container; Lyria 3.5 can also return WAV.

Multimodal input is supported. Up to ten images can accompany a text prompt, and the model composes against the visual content. For a marketer, that maps directly onto a workflow where a campaign key visual or a product photograph becomes the brief for the audio bed rather than a written description of it.

The prompting guidance published alongside the model reads like a production brief template: genre, named instruments, beats per minute, musical key, mood, structural tags such as [Verse] and [Chorus], and duration. Custom lyrics can be supplied inside the prompt using the same section tags. Timestamp control allows an instruction such as an intro from zero to ten seconds and a chorus from thirty to fifty, which is the mechanism that lets a generated cue land on a cut rather than near it.

Language handling follows the prompt. According to the documentation, lyrics are generated in the language the prompt is written in, and the model adapts vocal style and pronunciation accordingly. The response arrives as a sequence of steps containing separate text and audio blocks, with the generated lyrics and song structure returned alongside the base64-encoded audio.

The documented limitations are as consequential as the features. Music generation is single-turn: iterative refinement of a generated clip across multiple prompts is not supported in the current version. Results vary between calls even with an identical prompt. And safety filters block prompts requesting specific artist voices or the generation of copyrighted lyrics, a restriction that reads as a direct response to the litigation history of the category.

Provenance and the disclosure question

Every track carries a SynthID audio watermark. The documentation describes it as imperceptible to the human ear and states that it does not affect the listening experience.

SynthID is not new. DeepMind introduced it in 2024 as a suite of tools for marking and identifying machine-generated content across text, audio, images and video, designed to survive editing and reformatting. Google added consumer-facing verification to the Gemini app on December 18, 2025, allowing a user to upload a video and receive timestamp-specific feedback on whether audio, visuals or both carried the marker. When Gemini 3.5 Live Translate arrived in June 2026, its synthetic speech shipped with the same watermark woven into the audio rather than attached as separate metadata.

The regulatory context has hardened considerably since then. Article 50 of the EU AI Act became legally applicable on August 2, 2026. The European Commission published its implementing guidelines and a finalised Code of Practice on Transparency of AI-Generated Content on July 20, 2026, fixing the technical detail that providers and deployers of generative systems must meet. Providers face machine-readable marking duties; deployers, a category that expressly includes advertisers, agencies and publishers, face labelling duties where the content qualifies. Penalties reach 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.

An imperceptible watermark satisfies the provider-side marking obligation. It does not, by itself, satisfy a deployer-side labelling obligation, which contemplates a visible disclosure rather than an embedded signal. The same distinction already applies to a deepfake under Article 50(4), where the regulation calls for a visible label rather than the invisible marking permitted elsewhere. A brand jingle generated in the Gemini app and dropped into a paid social campaign inherits the marking but not the label.

Google has been building the advertiser-facing half of that stack separately. On July 9, 2026, the company introduced a How this ad was made disclosure panel across Google Ads, Display and Video 360, Campaign Manager 360, Merchant Center and Ads Editor, with automatic disclosure for assets produced in Google's own tools and manual controls for everything else. A track generated in the consumer Gemini app and uploaded by hand does not travel through that pipeline.

Where the model sits in Google's creative stack

Lyria has been threaded through Google properties for close to three years. YouTube ran Dream Track as a limited experiment from late 2023, generating short Shorts soundtracks in the voices of nine participating recording artists. In September 2025, the platform integrated Lyria 2 through a Speech to Song feature that converted dialogue from eligible videos into musical beds while preserving attribution to the original creator.

By mid-2026 the model had reached third-party advertising software. Typeface added Lyria-powered soundtrack generation to its Video Agent product in July 2026, with the system reading a video's duration and tone before producing a matching track, according to that company. Google and Typeface already held an integration relationship allowing creative assets to import directly into Google Ads.

YouTube's own creator surfaces now present generated audio as one option among three. When the platform extended background music to Shorts image posts in July 2026, it named a licensed and popular music library, the royalty-free YouTube Audio Library, and Dream Track, with the last carrying an explicit market-eligibility caveat the other two did not.

The distribution numbers explain why the Gemini app placement matters more than any of those earlier integrations. Alphabet reported 750 million monthly active users for the Gemini app in its fourth quarter 2025 results on February 4, 2026. The figure crossed 900 million by May 2026 and reached 950 million in the second quarter of 2026, with daily actives tripling year over year. Placing a music generator inside that installed base changes the volume question rather than the capability question.

Cost pressure runs in the same direction. On September 1, 2026, three days before this release, Google published benchmarks claiming token reductions of up to 88 percent for video analysis in agentic mode across three Gemini Flash models. The company has also been reworking the commercial guardrails around the API: Project Spend Caps arrived in Google AI Studio on March 16, 2026, seven months after a billing bug left some developers with charges running past 70,000 dollars for services they had not used.

The licensing environment this lands in

Generated music is not arriving into a settled legal picture.

On July 31, 2026, the 42nd Civil Chamber of the Munich I Regional Court ruled against Suno in a case brought by the German collecting society GEMA, setting a penalty of 250,000 euros per breach. The 143-page judgment ordered the defendant to stop reproducing six musical works for training purposes in the United States, to stop storing them inside the model, and to stop offering that model in Germany. The chamber applied United States copyright law directly and concluded that fair use did not shield the training.

Universal Music Group settled its infringement litigation with Udio in October 2025 and agreed licences for a walled-garden platform, a structure that an artist coalition challenged in November 2025 on transparency grounds. Rights holders have also begun applying sync-licence theory to ordinary brand social output: Sony Music filed a 58.8 million dollar claim against Kroger in August 2026 covering 392 songs used across the retailer's social accounts, citing 14 prior Sony licences since 2017.

That last case is the one that explains the commercial appeal of a model like Lyria 3.5 to a brand marketing team. The exposure in the Kroger complaint arises from using recorded music from a rights holder's back catalogue without a sync licence. A generated track carries no such counterparty. What it carries instead is uncertainty about the training data behind it, which is precisely the question the Munich court answered against Suno.

Platforms have meanwhile started building disclosure machinery on the distribution side. Apple Music introduced Transparency Tags on March 4, 2026, a metadata framework covering artwork, sound recordings, compositions and music videos, with the disclosure duty placed on labels and distributors rather than on the platform, and no published cross-verification mechanism.

Why this matters for the marketing community

Three practical consequences follow from the September 4 release.

The first is supply. A production music library, a stock music subscription and a bespoke composer commission all price against scarcity of finished audio. A model that returns a structured two-minute track with vocals from a text prompt, at no cost inside a consumer application reaching close to a billion monthly users, removes that scarcity for the low end of the market. The bespoke commission does not disappear. The 200-dollar library cue for a fifteen-second social edit has a harder case to make.

The second is compliance. The watermark answers a provider obligation. It does not answer the deployer obligation that falls on the advertiser or the agency, and it is invisible by design, which means no reviewer catches it by listening. Whether a generated track requires a visible label in a given campaign depends on the application, the legal role of the party publishing it and the content type, a scope question the Dutch industry association has recently worked through for creative agencies. Nothing in the Gemini app interface answers that question at the point of generation.

The third is verification. Google's own detection surface reaches only Google's own marker. A track produced in Lyria 3.5, exported, re-encoded and delivered to a supply-side platform is identifiable if a downstream system looks for SynthID and unidentifiable if it does not. Buyers have not, as a body, adopted that signal into brand safety classification, and the infrastructure for doing so exists on some platforms and not others.

The reasoning layer sits between the prompt and the audio. According to the documentation, Lyria 3.5 works through musical structure before generating, and the model uses an internal prompt rewriter that is not exposed to the user. Intermediate reasoning blocks and thought signatures are withheld. For anyone attempting to audit why a generated cue sounds like something familiar, that opacity is a structural limit rather than a configuration choice.

Timeline

Summary

Who: Google and Google DeepMind, which developed the Lyria model family. The audience spans consumer Gemini users worldwide, artists and creatives working in Google Flow Music, and developers building through Google AI Studio and the Gemini API. Marketers, agencies and publishers sit downstream as the parties who publish the output and carry the resulting disclosure duties.

What: Lyria 3.5, described by Google as its best-sounding music generation model, became available in the Gemini app and the Gemini API. The consumer interface gains genre selection or description, a vocal and instrumental switch, templates, and adjustable track length. The developer interface exposes two models: Lyria 3 Clip at a fixed 30 seconds, and Lyria 3.5 for songs lasting a couple of minutes with verses, choruses and bridges. Both return 44.1 kHz stereo audio, accept up to ten images alongside a text prompt, and embed a SynthID watermark in every output. Music generation is single-turn, results are non-deterministic, and safety filters block prompts requesting specific artist voices or copyrighted lyrics.

When: September 4, 2026 for the Gemini app and API announcement, with the Gemini API changelog listing the model in public preview from September 3, 2026, and an earlier Google Flow Music rollout on July 29, 2026.

Where: Globally, on the web and in the mobile Gemini application, and across Google Flow Music, Google AI Studio and Google Vids.

Why: Placing a music generator inside an application reporting 950 million monthly active users moves machine-composed audio from a specialist tool to a default one, at a moment when the licensing and disclosure environment around generated music is tightening. A Munich court quantified liability for unlicensed training at 250,000 euros per breach in July 2026, Article 50 of the EU AI Act began binding providers and deployers on August 2, 2026, and rights holders have started pursuing brands over recorded music in ordinary social posts. Generated audio removes the sync-licence counterparty and replaces it with a marking obligation that the watermark satisfies on the provider side and leaves open on the deployer side.