Summary: Google DeepMind released Lyria 3 Pro on March 25 as its most capable AI music model, composing tracks up to three minutes versus Lyria 3’s 30-second cap. It reached the Gemini app, Google Vids, Vertex AI, AI Studio, and ProducerAI, with lyria-3-pro-preview added to the Gemini API. Google says training used permitted material and every output carries SynthID watermarking.
Lyria 3 Pro arrived on March 25 — Google DeepMind’s most capable AI music generation model yet. Where Lyria 3, released just one month prior, topped out at 30 seconds per generation, the new model composes finished tracks running as long as three minutes. And the rollout was broad from day one: the model landed in the Gemini app, Google Vids, Vertex AI, AI Studio, and ProducerAI, the agentic music production tool Google bought in February. With that single release, AI-generated music stopped being an experimental snippet and reached lengths you can actually deliver.
What should catch a developer’s eye isn’t the composing itself. It’s the distribution: music generation now sits inside Google’s API surface and enterprise channels. A public preview on Vertex AI plus model access through the Gemini API means audio generation is being packaged as a billable, scalable primitive — held at the same tier as text and images.
Structured Composition, Not Stitched Clips
The extra length is not the result of gluing shorter pieces together. Per Google, the model understands how a song is assembled at a deeper level: name the building blocks you want — intros, verses, choruses, bridges — directly in the prompt, and Lyria 3 Pro takes care of moving between sections and experimenting across styles. TechCrunch called the leap beyond the previous generation’s 30-second cap a major one.
In practice, the difference changes what a single generation can cover. Thirty seconds buys you a demo or a sample; three minutes is enough for a short video, a podcast opening, or scoring a scene inside a game — game scene scoring being the very use case Google called out in the Vertex AI preview. Working musicians also helped shape the model through the Music AI Sandbox: Grammy winner Yung Spielburg scored the DeepMind short film “Dear Upstairs Neighbors” with Lyria, and DJ and producer François K used it for an upcoming release.
Where the Model Landed: Products and APIs
Every surface the update covers is worth spelling out:
- Vertex AI: a public preview for businesses that need audio generated on demand and at scale
- AI Studio and the Gemini API: developers can combine Lyria RealTime with it to build real-time creative tools
- Google Vids: beginning to roll out that week to Workspace customers and Google AI Pro and Ultra subscribers
- Gemini app: paid subscribers get longer generations, with vlogs, podcasts, and tutorials named as intended uses
- ProducerAI: the agentic music production tool acquired in February, available worldwide to free and paying users
That same day, lyria-3-pro-preview showed up in the Gemini API changelog, confirming the API route opened in step with the product launches. That products-first, API-close-behind rhythm matches the model acceleration we followed in our opening-of-2026 outlook — the difference being that audio is what speeds up this time.
Training Data and How the Licensing Line Is Drawn
The data question catches up with every music model. Google’s position: training drew only on material that YouTube and Google have a right to use under their terms of service, partner agreements, and applicable law. The models “do not mimic artists” — a creator’s name in a prompt serves purely as broad stylistic inspiration — and every output passes through filters that compare it against existing content. Each track from Lyria 3 and Lyria 3 Pro also carries a SynthID watermark marking it as AI-generated.
Set against the wider industry, the caution looks engineered rather than incidental. Spotify already offers artists tools to review songs released under their name, and Deezer built detection tooling that other streaming services can use to flag AI music. Google’s three-part bundle — permitted-or-licensed material, watermarking, no mimicry — amounts to a strategy for steering wide of the litigation territory where Suno and Udio have spent the past two years.
The Takeaways for Developers and Creators
Three things follow. One: once audio generation is exposed as an API, the cost math becomes comparable, and the “good enough” tier of audio — game soundtracks, video scoring, podcast beds — is the first to get priced against generation costs. Two: SynthID arriving as a factory default turns detection and disclosure of AI audio into platform-level infrastructure rather than a policy each company sets on its own. Three: throwing ProducerAI open to free users worldwide is customer acquisition and moat-building at once — once a producer’s sample library and workflow settle inside Google’s ecosystem, the cost of leaving only grows.