Lyria 3.5 First Look: Inside Google DeepMind's Newest AI Music Model
Some links in this guide are affiliate links. If you sign up through them, Day Zero Guides may earn a commission at no extra cost to you. This never affects which products we cover or what we say about them. See our affiliate disclosure for details.
Google DeepMind quietly pushed out Lyria 3.5, the latest revision of its music generation model, and it's worth a close look — not because Google has published a slick self-serve pricing page (it hasn't), but because Lyria has quietly become the engine behind more shipped products than any other AI music model on the market. If you've used YouTube's Dream Track feature, played with MusicFX or MusicFX DJ in Google Labs, or tried the real-time jamming demos DeepMind showed off at I/O, you've already heard a Lyria model at work. Version 3.5 is the next step in that lineage, and it's aimed squarely at closing the audio-quality and controllability gap with tools like Suno and Udio rather than chasing academic novelty.
What Lyria 3.5 actually does
Lyria 3.5 generates original musical audio — not MIDI, not stems assembled from a sample library, but full waveform audio — from text prompts describing genre, mood, instrumentation, and structure. DeepMind's public materials emphasize three things carried forward and sharpened from Lyria 2: higher-fidelity output (44.1kHz-class audio rather than the muddier, lower-bitrate output that plagued earlier generative audio models), tighter prompt adherence for genre and instrumentation, and support for the "RealTime" interactive mode that lets a model respond to live parameter changes (tempo, intensity, instrument mix) the way a human DJ or session musician would react to a bandleader's cues, rather than requiring a full re-generation for every tweak.
This matters because most AI music tools force you into a one-shot workflow: type a prompt, wait, get a finished clip, repeat if you don't like it. Lyria's RealTime approach is closer to a live instrument — you can nudge a generation while it's playing, which is a meaningfully different product experience than a batch text-to-music generator.
Pricing: not published, and that's the catch
As of this writing, Google has not released standalone consumer pricing for Lyria 3.5. There is no public "$X/month" tier, no credit-based checkout, and no self-serve API key you can generate today the way you can with Stability AI or Meta's open weights. Access is currently gated in one of two ways:
- Through Google products that embed Lyria, like MusicFX and MusicFX DJ in Google Labs (free, waitlist/region-gated at times) and YouTube's Dream Track (rolled out to a limited creator set).
- Through Vertex AI for enterprise customers, where prior Lyria versions have been billed on Google Cloud's standard usage-based model (charged per second of generated audio, billed alongside your other Vertex AI model usage) rather than a flat subscription.
If you're evaluating Lyria 3.5 today, budget for the likelihood that broad public access will follow the same pattern: a free, feature-limited consumer surface (Labs/MusicFX) and a metered enterprise API (Vertex AI) rather than a $20/month indie-creator tier like Suno or Udio offer. That's a real limitation if you're a solo musician or small studio expecting to buy a seat and start exporting tracks tonight.
Concrete use cases
- Short-form video scoring at scale: YouTube's Dream Track integration is the clearest signal of intent — creators type a prompt, get a needle-drop-style backing track, and drop it into a Short without licensing a sample or hiring a composer.
- Interactive/live audio experiences: The RealTime mode is being pitched at game studios and live-performance/installation artists who need music that shifts continuously with user input (game state, crowd energy, sensor data) instead of looping pre-rendered stems.
- Enterprise content pipelines: Vertex AI access targets media companies that want programmatic background-music generation for ad reads, podcasts, or in-house video at volumes where licensing stock music per-clip gets expensive.
- Rapid ideation for composers: Feeding in a mood and instrumentation brief to get a sketch or motif to react to and rework in a DAW, rather than shipping the raw output as a finished master.
What it's not yet clearly positioned for: bedroom producers who want full-song generation with vocals and a downloadable stem pack for $10/month — that's still Suno/Udio territory, and Google hasn't matched their self-serve packaging.
How it compares
| Lyria 3.5 (Google DeepMind) | OpenAI Jukebox | Stable Audio (Stability AI) | MusicGen (Meta) | |
|---|---|---|---|---|
| Price | Not published; free via Labs/MusicFX (limited access), metered via Vertex AI for enterprise | Free, open-source research release — no hosted product or ongoing maintenance | Free tier limited; Stable Audio Pro subscription ~$11–12/month, Studio tier for commercial rights | Free and open-source (weights on GitHub/Hugging Face); no official hosted paid tier |
| Output quality/speed | High-fidelity audio, includes real-time/interactive generation mode | Raw audio with vocals, but notoriously slow (historically took hours to render minutes of audio) | Fast generation, up to ~3 minutes of audio per clip, tuned for production use | Fast generation for instrumental clips; quality noticeably behind commercial tools |
| Vocals support | Emphasis on instrumental and structured music; vocal handling not a headline feature yet | Included (its signature feature) but low fidelity by modern standards | Instrumental-focused | Instrumental only |
| Access model | Gated: consumer via Google Labs products, enterprise via Vertex AI | Fully open, self-hosted, no support or updates | Self-serve web app and API with commercial licensing tiers | Self-hosted only; requires your own GPU infrastructure and ML familiarity |
| Best for | Teams already in the Google ecosystem (YouTube creators, Vertex AI customers) wanting interactive or embedded music generation | Researchers studying early generative audio architecture; not a practical production tool today | Indie creators and small studios wanting a straightforward paid tool with commercial rights today | Developers who want a free, customizable, self-hosted model and don't mind running inference themselves |
Should you try it today?
If you're a solo creator hunting for a $10-a-month tool to generate finished, licensable tracks right now, Lyria 3.5 isn't that yet — Stable Audio's Pro tier is the more turnkey option, and it has a checkout page. If you're already inside the Google Cloud ecosystem, building a product that needs programmatic or interactive music generation, or you're a YouTube creator with Dream Track access, Lyria 3.5 is worth requesting access to immediately — the RealTime interactive mode in particular isn't something Stable Audio or MusicGen currently offer in a comparable form. Everyone else should bookmark it and check back once Google publishes actual consumer pricing; judging by how Lyria 2 was handled, that could arrive quietly inside an existing Google product rather than as a standalone launch.