Google Makes Lyria 3.5 AI Music Available in Gemini App and Developer API

On September 4, Google made Lyria 3.5 — its current flagship AI music generation model — available to every Gemini user in the world and to any software developer who already holds a Gemini API key. The same model had quietly launched in the specialist creator platform Google Flow Music on July 29, where it sat largely unreachable by general audiences for five weeks. The September 4 rollout is not a model upgrade; it is a distribution event, and the distinction matters: a music generation capability that most producers had to actively seek out is now the default option inside the world's most widely used AI chatbot, wired behind the same API surface developers already use for Gemini text and code.

Staged Rollout Reveals Google's Product Strategy

The announcement was framed as a single product launch, but the public record tells a more layered story. Lyria 3.5 first appeared in Google Flow Music on July 29, when Google described improvements to musicality, lyrics, vocals, and creative control over the Lyria 3 Pro model that had launched in March. The Gemini API changelog listed Lyria 3.5 in public preview on September 3, one day before the consumer announcement. The September 4 post described the model as "now available" without acknowledging the prior Flow Music release — a presentation choice that led most coverage to treat the announcement as a new model rather than an access expansion.

This staged approach reflects a deliberate product strategy. Google first validated Lyria 3.5 with the smaller audience of active Flow Music creators — a population already engaged enough with AI music tools to seek out a specialized platform — before opening the same model to the full Gemini user base and the developer community simultaneously. The result is that Lyria 3.5 enters its widest deployment window with at least six weeks of real-world usage behind it.

Read more: Google Folds Speech-to-Text Into Gemini API as Transcribe Enters Public Preview

How the Lyria Stack Actually Generates Music

Lyria 3.5 is not a single model. The underlying architecture that powers it inside both Flow Music and the Gemini app is a multi-model stack: Lyria itself handles audio generation, Gemini provides the natural language understanding that interprets a user's creative prompt, and Veo handles visual components when the broader Flow Music ecosystem is involved. This division of labor is significant because it means improvements to Gemini's reasoning capabilities — which occur independently of Lyria updates — can improve prompt interpretation for music generation even when the audio model itself hasn't changed.

The audio output is produced at 44.1 kHz in stereo, which matches professional recording standards. According to Google's developer documentation, developers accessing the model through the Gemini API can choose between two model variants: lyria-3.5, which generates full-length compositions with verses, choruses, and bridges lasting up to three minutes, and lyria-3-clip-preview, which always produces 30-second clips designed for loops and audio previews. Both accept text prompts and — uniquely — images. Users can supply up to ten images alongside a text description, and Lyria 3.5 will compose music inspired by the visual content.

What separates the current generation from earlier AI music systems is song-structure awareness. Earlier approaches treated audio generation as a token-by-token continuation problem — the model produced sound without knowing whether it was in an intro, a verse, or a bridge, which produced extended loops rather than compositions. Lyria 3.5 was trained to understand songs as structured entities with distinct sections that serve different narrative and emotional functions. Developers can now provide explicit section tags in their prompts — marking Verse, Chorus, and Bridge — and add timestamp-based triggers to specify when particular instruments or events should enter a track. The model adjusts vocal style and pronunciation to match the language of the prompt automatically. Before generating audio, the system reasons through the musical structure implied by the prompt, which Google says produces stronger structural coherence in the output.

SynthID Has Become the Industry's Provenance Standard

Every audio file Lyria 3.5 generates carries a SynthID watermark. The mechanics of SynthID are worth understanding in detail, because the technology has grown well beyond its origins as a Google product feature.

SynthID uses psychoacoustic masking to embed a cryptographic signature directly into the audio waveform at generation time. Psychoacoustic masking is the phenomenon by which certain sounds become inaudible when other sounds are present — exploiting this, SynthID places its signal in frequency ranges where the human ear is least sensitive and during segments where louder audio naturally masks quieter signals. The result is a watermark that is structurally part of the audio itself, not a detachable metadata tag. Documented tests show the signature survives MP3 compression — including aggressive 128 kbps encoding that strips conventional metadata — as well as speed and pitch manipulation, and even re-recording a track through a physical speaker into a new microphone.

The adoption curve for SynthID has accelerated sharply. In May 2026, OpenAI incorporated SynthID into images generated by ChatGPT, and NVIDIA announced integration into its Cosmos world models. These are competing organizations adopting a Google-originated standard, which rarely happens in AI unless the technical need is acute — and in this case it is, because streaming platforms have implemented mandatory AI-content disclosure policies that require automated detection infrastructure. Spotify's AI Credits disclosure framework launched on April 16, 2026; TIDAL's AI labeling policy took effect July 15, 2026; Deezer has deployed an AI-detection system it claims achieves 99.8% accuracy. SynthID is the mechanism by which Lyria-generated audio gets identified reliably and automatically by these systems. Neither Suno nor Udio has adopted SynthID.

For creators who intend to distribute Lyria-generated music through streaming platforms or embed it in commercial products in Europe, where the EU AI Act's transparency requirements for AI-generated content took effect in August 2026, SynthID's chain of provenance is not incidental — it is the compliance mechanism.

How Lyria 3.5 Compares to Suno, Udio, and ElevenLabs

Audio quality comparisons among AI music generators are contested and fast-moving, but the current landscape has a relatively clear structure.

Suno's most recent models hold the quality leadership position on available benchmarks. When Suno launched its v5 model in late 2025, the company's own internal ELO benchmark placed it at a score of 1,293 — ahead of prior Suno versions and, by the company's assessment, ahead of every competitor on audio fidelity, musical structure, and vocal realism. Suno has since released v5.5 (March 26, 2026) as its current stable model. Suno's vocal range — capturing whispers, vibrato, breathiness, and emotional nuance — remains the capability Lyria 3.5 is explicitly targeting with its improvements to expressiveness and pronunciation. Suno has reported 2 million paid subscribers and $300 million in annual recurring revenue as of February 2026, a $5.4 billion Series D valuation as of June 2026, and over 100 million total users since launch. Despite this commercial momentum, Suno has not launched an official public self-serve API; developers who want programmatic access rely on third-party wrappers, which charge approximately $0.05 to $0.11 per generation.

Udio occupies a distinct niche. Built by former Google DeepMind researchers, Udio has developed a reputation for strong performance specifically in electronic music, hip-hop, and pop — genres where timing precision and production crispness matter most — as well as orchestral generation, where its output has depth and spatial quality that community reviewers consistently rate above competitors. Udio also offers stem separation, giving producers individual instrument tracks to work with downstream.

ElevenLabs Music represents the most developer-accessible alternative on the commercial licensing side. Its Music API is the most mature in the field for programmatic integration and carries explicit commercial use rights. The tradeoff is cost: ElevenLabs charges approximately $0.30 per minute of generated audio, translating to roughly $0.60 to $0.90 for a two-to-three minute track — substantially higher than the pricing structure available through Google's Gemini API, which is bundled with existing developer access rather than priced separately per generation.

Google's structural position is unusual. On audio quality, Lyria 3.5 trails the Suno model line as measured by available benchmarks and lags Udio's community reputation in the genres where Udio specializes. On distribution, however, Google has no peer. The Gemini user base reaches a general audience that dwarfs any music-specific platform; the API is available to any developer who already has a Google account. Google's licensed-data claim and SynthID watermarking give it the cleanest commercial posture in the space. The September 4 rollout is the moment when Google's distribution advantage becomes real rather than theoretical.

Documented Limitations and Commercial Deployment Caveats

The developer documentation for Lyria 3.5 explicitly describes several constraints that commercial buyers need to account for. Generation is a single-turn process: there is no iterative editing of a generated clip within the current API version. If a first result misses the mark, the workflow is to refine the prompt and generate again rather than modify the output. All prompts pass through safety filters; requests that ask for a specific named artist's voice, or that submit copyrighted lyrics for direct reproduction, are blocked by the model. Results vary between calls even with identical prompts, which has implications for any workflow that requires deterministic or consistent output.

The training data question remains the most consequential uncertainty for commercial deployment. Google characterizes Lyria 3.5 as trained on licensed content — materials that YouTube and Google have rights to use under their terms of service, partner agreements, and applicable law. That claim is company-stated and has not been independently audited. In parallel, a group of independent musicians and songwriters filed a copyright lawsuit in March 2026 alleging that Google used approximately 44 million audio clips totaling roughly 280,000 hours of copyrighted recordings sourced from YouTube to train Lyria 3 — the direct predecessor to Lyria 3.5 — without authorization. Google filed a motion to dismiss in June 2026, arguing that YouTube's terms of service constitute a binding license for this use. The case remains active. Developers building commercial products on Lyria 3.5 output should verify their specific use cases against Google's published commercial licensing terms before launch rather than relying on the general "licensed data" characterization.

Lyria's Place in a Market Being Reshaped by Legal Risk

The copyright litigation surrounding AI music is more advanced than most general-purpose AI copyright disputes, and the outcomes matter directly to anyone choosing an AI music platform.

The Recording Industry Association of America filed copyright suits against Suno and Udio in June 2024 on behalf of major label plaintiffs. Universal Music Group reached a settlement with Udio in October 2025; Warner Music Group reached a settlement with Suno in November 2025. Sony Music Entertainment, however, has not settled with either company, and its cases remain fully active. In July 2026, an analysis of hacked Suno source code reported by 404 Media and independently confirmed by TechCrunch and Music Business Worldwide showed that the company had used a proxy tool to stream-rip audio from YouTube Music, Deezer, and other streaming services for training data. Suno has argued that training on music available on the open internet is protected as fair use under applicable law, but the confirmed evidence of stream-ripping created material reputational and legal exposure that remains unresolved.

This litigation landscape creates asymmetric risk for commercial music deployments. For a brand using AI music in advertising, a creator licensing tracks for film or television sync, or an enterprise embedding background music generation in a product — the identity of the training data and its legal provenance is not an abstract question. Google's SynthID watermarking and its licensed-data framing position Lyria 3.5 as the lower-risk option among the field's highest-profile tools, even as the indie artist lawsuit creates residual uncertainty about that characterization. ElevenLabs Music offers a comparable legal posture with a more mature API but at significantly higher per-generation cost.

The competitive question is whether that structural safety advantage is sufficient to redirect workflows that have already formed around Suno's quality leadership or Udio's genre specialization. For new deployments — particularly enterprise applications and anything touching streaming platform distribution — the Lyria 3.5 rollout into a public API creates a viable path that didn't exist before September 4. For producers with existing Suno or Udio workflows, the calculus is more friction-dependent, and Google's audio quality will need to close the remaining gap to generate meaningful switching.

What the September Rollout Changes for Developers

The most underreported aspect of the September 4 announcement is what it means for developers specifically. Lyria 3.5 is now accessible via the same Gemini API key developers already hold for text, code, and reasoning tasks. There is no separate music API account to create, no separate waitlist, and no new billing relationship to establish. A developer building a video production tool, a game, an interactive application, or a podcast platform can wire in music generation through the same API surface they already use — a single-session integration path that ElevenLabs and third-party Suno wrappers cannot replicate.

The pricing contrast with competitors is significant. ElevenLabs Music, the market's most mature commercial audio API, charges approximately $0.30 per minute of generated audio — meaning a two-to-three minute track costs roughly $0.60 to $0.90. Third-party API wrappers for Suno, which has no official public API of its own as of this writing, charge between $0.05 and $0.11 per generation. Lyria 3.5 via the Gemini API is priced on standard Gemini API usage tiers rather than as a per-track charge, which makes per-unit costs dependent on a developer's existing plan and usage volume. For organizations already inside the Google Cloud and Gemini API ecosystem, the marginal cost of adding music generation may be substantially lower than switching to a dedicated music API. This economic structure — bundling music with text and reasoning rather than pricing it separately — is a differentiation strategy no standalone music AI company can replicate.

The Gemini API music generation documentation is currently in public preview, which means its commercial terms and production-readiness guarantees are not finalized. Google AI Studio offers a prompt testing environment where developers can iterate on music generation prompts before writing API integration code — the same prompts transfer to the production API with minimal rework. Organizations evaluating Lyria 3.5 for production deployment should treat the current moment as an evaluation window rather than a production-ready launch, confirm specific use cases against Google's published licensing terms, and monitor for when the model exits public preview status.

Separately, the September 4 rollout coincided with Google consolidating speech-to-text capabilities under the Gemini API through the Transcribe feature entering public preview. The parallel suggests a broader strategy: Google is systematically routing modality-specific AI capabilities — text generation, image generation, video generation, music generation, and now speech transcription — through the unified Gemini API surface. For developers, this reduces integration overhead and creates a single dependency. For competitors who built standalone audio and music APIs, it increases the switching cost of moving to their platform, because staying with Lyria means no additional integration work.

The measure of whether Lyria 3.5's distribution advantage translates into developer adoption is how quickly independent tools built on the Gemini API emerge over the coming months, and whether the audio quality improvements Google describes are sufficient to satisfy production requirements. Suno's established workflows and model line give it staying power; Udio's genre-specific strengths give it protection in its niches. What Google's September rollout adds is a viable licensed-data, SynthID-watermarked, single-API option at a price point that makes evaluation frictionless — and that is a structural shift in the market even before the next model generation arrives.