Lyria 3.5 in Gemini 2026: Make Studio Tracks Free

Google shipped Lyria 3.5 into the Gemini app while everyone argued about GPT-6 Astra, and almost nobody noticed that a frontier lab just handed out a studio-grade music engine for free. Lyria 3.5 Gemini is the first time full song generation — structured arrangements, separated stems, and generated vocals — lives inside a mainstream chat assistant rather than behind a specialist product like Suno or Udio. Open the Gemini app right now, on a free tier, type a paragraph describing a track, and you get back something that survives a listen on real speakers. That combination of capability and distribution is the story, and it is badly underrated.

What’s actually new in Lyria 3.5 Gemini

Lyria has existed in some form since the DeepMind music experiments and the earlier Lyria 2 model that powered MusicFX and the Music AI Sandbox. Those were demos with a hard ceiling: roughly 30 to 60 seconds of coherent audio, instrumental-only output on most surfaces, and no way to revise one part of a track without regenerating the whole thing. Lyria 3.5 breaks all three limits at once. Structural coherence now extends across multi-minute pieces, so the model holds a key, keeps a motif, and returns to a chorus instead of drifting into ambient mush around the 45-second mark.

The bigger technical shift is stem-level control. Lyria 3.5 can emit drums, bass, harmonic instruments, and vocals as separate tracks rather than a single baked-down mix. That is the difference between a novelty and a tool. Stems let you drop a generated bassline into a DAW, mute the AI drums in favor of your own, or hand a producer something they can mix. Vocal generation arrives alongside it — the model sings lyrics with controllable timbre and delivery, closing the single biggest gap between Google’s music work and what Suno shipped two years ago.

Distribution is the third piece. Google Gemini music AI is now a first-class capability inside the assistant most Android phones ship with, reachable from the same text box you use for email drafts. Free-tier users get generation with rate limits; Google AI Pro and Ultra subscribers get longer outputs, faster queues, and stem downloads. Every clip carries a SynthID music watermark — an inaudible signal embedded in the waveform that survives compression, EQ, and re-encoding, and that Google’s verification tooling can detect after the fact.

Why it matters

• The price floor for background music just hit zero. Anyone paying a monthly stock-music subscription for podcast beds, YouTube intros, or ad underscores now has a free alternative that produces bespoke output instead of a library search.
• Stems change the workflow from “generate” to “produce.” Once you can isolate parts, AI audio stops being a finished artifact you accept or reject and becomes raw material you edit — the same shift that made image models useful to designers rather than a curiosity.
• Suno and Udio lose their moat on distribution, not quality. A dedicated app has to win a user’s attention. Gemini already has it. Even if Lyria 3.5 vs Suno lands as a rough tie on audio quality, “already installed” is a brutal advantage.
• Watermarking becomes the default. SynthID coverage on every output gives platforms, labels, and rights holders a detection path that does not depend on the uploader being honest.
• Prompt-to-track collapses a whole freelance tier. The $50-to-$300 custom jingle and podcast-theme market is the most exposed. Higher-end scoring work is safer, but the bottom of that market is being priced out in real time.
• Multimodal assistants now cover text, image, video, and audio. Lyria 3.5 was the last missing modality in Gemini. The next round of feature work will chain them — a Veo video that scores itself.

How to use Lyria 3.5 in the Gemini app today

1. Open Gemini and ask for music explicitly. Most accounts have no separate mode toggle; the router picks Lyria when your prompt is clearly musical. Being explicit keeps you from getting a text description of a song instead of a song. Good prompts name genre, tempo, instrumentation, mood, and structure:

   Generate a 2-minute instrumental track with Lyria.
   Genre: downtempo electronic with live-feel drums
   Tempo: 92 BPM, 4/4
   Key: F minor
   Instrumentation: Rhodes electric piano, sub bass, brushed
     drum kit, tape-saturated pads
   Structure: 8-bar intro, 16-bar A section, 8-bar breakdown
     with drums dropped out, 16-bar A section with added
     arpeggio, 8-bar outro that filters down
   Mood: reflective, late-night, not sad
   Mix: warm, slight vinyl noise floor, no sidechain pumping

2. Add vocals only when you have actual lyrics. Vocal generation degrades fast on vague prompts. Give it the lines and describe the voice as you would to a session singer:

   Same track, now add a lead vocal.
   Voice: female alto, breathy, close-mic'd, minimal vibrato
   Delivery: conversational on verses, sustained on the hook
   Lyrics:
   [Verse]
   Neon on the wet glass, engine idling low
   Every road out of here goes the way I already know
   [Hook]
   So hold the line, hold the line
   I'll be home before the streetlights die

3. Revise one element instead of rerolling. The whole point of the upgrade is that you can address parts. Reference the section and the instrument:

   Keep everything. Two changes only:
   1. In the breakdown, replace the pad with a solo cello
      playing long tones under the Rhodes.
   2. Make the outro four bars longer and end on a single
      sustained F minor chord, no drums.

4. Export stems for anything you plan to publish. Ask for separated tracks rather than the mix. Pro and Ultra tiers return individual WAV files; free may return a mix plus a limited split:

   Export this as separated stems:
   drums, bass, keys, pads, lead vocal, backing vocals.
   WAV, 48kHz, 24-bit, no master bus compression or limiting.

5. Verify the SynthID watermark before you ship. If you are handing a client a track, confirm what it is. Google’s SynthID Detector accepts audio uploads and reports whether a Google model produced it — run your own file through it so you know how it will read to anyone else who checks.

6. For programmatic use, go through the Gemini API rather than the app. The music endpoints follow the same shape as the rest of the SDK:

   from google import genai
   
   client = genai.Client(api_key="YOUR_KEY")
   
   response = client.models.generate_content(
       model="lyria-3.5",
       contents=(
           "Instrumental cinematic build, 110 BPM, D minor. "
           "Strings enter at bar 9, taiko drums at bar 17, "
           "full ensemble at bar 25, hard stop at bar 33."
       ),
       config={
           "duration_seconds": 90,
           "output_format": "wav",
           "return_stems": True,
       },
   )
   
   for stem in response.audio_stems:
       with open(f"{stem.name}.wav", "wb") as f:
           f.write(stem.data)

   Check the current model card before wiring this into anything real. Parameter names on new audio models move around in the first few months, and free-tier API quotas for audio are tighter than for text.

7. Finish in a DAW. Import stems into Reaper, Ableton, or even Audacity, trim the intro to the length you need, and run a light limiter. Generated audio sits a few dB quieter and less dense than commercial reference tracks; ten minutes of mixing closes most of that gap.

How it compares

Capability	Lyria 3.5 (Gemini)	Suno	Udio	Stable Audio
Free tier for full songs	Yes, rate-limited	Limited daily credits	Limited daily credits	Limited, short clips
Vocal generation	Yes	Yes, strongest in class	Yes	Weak
Stem separation on export	Yes, paid tiers full set	Yes, paid	Yes, paid	Partial
Long-form coherence	Multi-minute	Multi-minute	Multi-minute	Shorter, loop-oriented
Section-level revision	Conversational, in-chat	Extend and replace tools	Inpainting-style edits	Limited
Watermarking	SynthID on all output	Metadata-based	Metadata-based	Metadata-based
Lives inside a general assistant	Yes	No	No	No
Commercial use	Tier-dependent, check terms	Paid plans	Paid plans	Paid plans

The honest read: Suno still has the edge on vocal realism and on the “sounds like a record” quality that comes from being a focused product with a tight feedback loop. Lyria 3.5 wins on instrumental control, on stem cleanliness, and on costing nothing and requiring no new account. For an AI song generator shortlist in 2026, that puts Lyria first for anyone making beds, scores, and underscores, and Suno first for anyone trying to make a song someone would stream.

What’s next

The obvious roadmap item is cross-modal chaining. Google already has Veo for video and Gemini for scripts; the missing piece was a scoring engine that understood timing. Expect scoring-to-picture — feed a video, get a track that hits your cuts — to ship as a feature rather than a research demo. That is where the real professional value sits, because syncing music to edits eats a video editor’s afternoon.

Watch the licensing fight. Google has led with SynthID and with statements about training data partnerships, which reads as a preemptive move against the litigation that hit Suno and Udio. A credible claim to a licensed or cleared training corpus turns a compliance headache into a selling point for exactly the commercial customers — agencies, broadcasters, brands — who cannot risk a model with murky provenance. That, not audio quality, will likely decide who wins the professional segment.

On the technical side, watch real-time generation latency, MIDI export alongside audio stems, and any form of style conditioning from a reference track. MIDI would be the single biggest upgrade for working musicians, because it makes the output fully editable rather than merely separable. Keep an eye on the free tier, too: Google launches generous, then tightens once the surface is popular. If Lyria 3.5 in Gemini matters to your workflow, build the habit now and export what you need.

Frequently Asked Questions

Is Lyria 3.5 actually free?

Yes, with limits. Free Gemini accounts generate music under rate caps and shorter maximum durations. Google AI Pro and Ultra raise those caps and unlock the full stem export set and faster queues. There is no separate music subscription.

Can I use Lyria 3.5 tracks commercially?

It depends on your tier and Google’s current terms, which you should read before invoicing anyone. Google has historically restricted commercial use on free consumer tiers and permitted it on paid plans. Check the terms attached to your account, and keep a record of the date you generated the track.

How does the SynthID music watermark work, and can I remove it?

SynthID embeds an inaudible signal directly into the audio waveform, so it survives MP3 compression, EQ, pitch shifting, and re-recording in ways that file metadata does not. Do not try to strip it — it exists so platforms can identify AI-generated audio, and attempting removal likely violates the terms you agreed to. Google’s SynthID Detector will tell you whether a given file carries one.

Lyria 3.5 vs Suno — which should I actually use?

Use Lyria for instrumentals, background beds, and anything where you want to control arrangement and edit stems afterward. Use Suno when you want a finished song with convincing vocals and you care more about the result than the process. Many people will use both, and trying Lyria first costs nothing since it is already in your Gemini app.

Does it work with my DAW?

Indirectly. Lyria 3.5 exports audio files, not project files, so the workflow is generate, download stems, import as audio tracks. There is no plugin and no MIDI export yet. That covers most editing work, but you cannot change the notes the model played — only how they sound and where they sit.

What are the biggest current limitations?

Vocal realism still trails the best dedicated tools, especially on complex phrasing. Long tracks can lose energy in the back half. The model refuses prompts that name specific living artists or attempt voice cloning. And free-tier generation queues slow down at peak times, which matters if you are iterating on a deadline.