Google Lyria Explained: Every Model, Spec, and Price from Lyria-002 to 3.5 (August 2026)
Last updated August 7, 2026
Google Lyria is DeepMind's music family: lyria-002 (30s instrumental, 48 kHz), Lyria 3 Clip (30s, vocals + timed lyrics) and Pro (184s), and Lyria 3.5 (up to 3 minutes). Every output carries a SynthID watermark; generation is single-turn with no iterative editing. Official pricing: $0.04 per Clip song, $0.08 per Pro song. All three picker variants run inside the invideo agent's Music tab.
Updated August 2026
Google Lyria is the provenance-first pick among AI music generators: every output from Google DeepMind's music family carries an imperceptible SynthID audio watermark, and guardrails block artist imitation outright — a stance that counts for more than any spec while the music-AI lawsuits below remain unresolved. The family is served through Vertex AI and the Gemini API and is available inside the invideo agent. As of August 2026 the lineup runs from lyria-002 (30-second, 48 kHz instrumental-focused WAV output) through the Lyria 3 pair — Clip for 30-second songs with vocals and timed lyrics, Pro for tracks up to 184 seconds — to Lyria 3.5, with improved vocals and song lengths up to 3 minutes. Official per-song pricing is $0.04 (Lyria 3 Clip) and $0.08 (Lyria 3 Pro).
How did Google Lyria evolve from lyria-002 to Lyria 3.5?
The family has moved fast, and most coverage conflates its generations. The dated record, from Google's own release notes and model pages:
- May 20, 2025 — Lyria 2, model ID
lyria-002, reaches general availability on Vertex AI, per Google's Vertex AI release notes. - March 25, 2026 — Lyria 3 enters public preview in two variants:
lyria-3-clip-preview(up to 30 seconds) andlyria-3-pro-preview(up to 184 seconds), per the same release notes. - 2026 — DeepMind's Lyria model page lists Lyria 3.5 as "our latest model with improved vocals, musicality, and variable song lengths," generating anything from 60-second clips to full 3-minute songs. A Vertex model ID for 3.5 had not been published at the time of writing.
- Lyria RealTime is a separate model for interactive, streaming music generation — not covered here.
Which Lyria model should you pick?
Inside invideo's model picker, the family appears as three entries. Here is what each maps to:
| invideo picker entry | Model | Max length | Vocals | Best for |
|---|---|---|---|---|
| Google Lyria 3 Pro | lyria-3-pro-preview |
184 seconds | Yes, with timed lyrics | Full songs, soundtrack beds for longer videos |
| Lyria 3 Clip | lyria-3-clip-preview |
30 seconds | Yes, with timed lyrics | Intros, ads, shorts, hook-length cues |
| Lyria-002 | lyria-002 |
30 seconds | Instrumental-focused | Clean instrumental beds at 48 kHz |
The length column decides most of this: only Pro clears 30 seconds, so anything longer than a hook is a Pro job by default.
What are Google Lyria's specs?
| Spec | lyria-002 | Lyria 3 Clip | Lyria 3 Pro |
|---|---|---|---|
| Duration | 30 s | up to 30 s | up to 184 s |
| Sample rate | 48 kHz | 44.1 kHz stereo | 44.1 kHz stereo |
| Format | WAV | MP3 | MP3 or WAV |
| Vocals + lyrics | Instrumental-focused | Yes | Yes |
| Watermark | SynthID | SynthID | SynthID |
| Status (Aug 2026) | GA | Preview | Preview |
One spec is usually reported backwards: lyria-002 outputs 48 kHz audio, while the newer Lyria 3 models are documented at 44.1 kHz high-fidelity stereo in the Gemini API music docs. Newer does not mean a higher sample rate here; 44.1 kHz is the CD standard, but the split is real and official.
Lyria 3 also takes multimodal input — a text prompt plus up to 10 images to condition the music — writes lyrics "in the language of your prompt," and accepts custom song structure via timestamps and section tags, all per the Gemini API documentation.
What does Google Lyria cost?
Official numbers only: the Gemini API pricing page lists Lyria 3 Clip Preview at $0.04 per song and Lyria 3 Pro Preview at $0.08 per song as of August 2026. Google has not published a comparable per-song figure for Vertex-side lyria-002 that we could verify, so we won't quote one.
What are Lyria's documented strengths and limitations?
Strengths. Provenance is the headline: DeepMind states that "all of our tracks are imperceptibly watermarked with SynthID." Lyria 3's timed-lyrics and section-tag control is unusual among music models, image-conditioned generation is distinctive, and the family ships through Google's enterprise stack rather than a hobbyist app.
Limitations, from Google's own docs. The Gemini API documentation is unambiguous about the biggest one: "Music generation is a single-turn process. Iterative editing… is not supported." You cannot ask Lyria to revise a track — you regenerate with a better prompt. Results are non-deterministic, prompts requesting a specific artist's voice or copyrighted lyrics are blocked by guardrail, and both Lyria 3 models remain in Preview as of August 2026, which Google notes may mean changes and tighter rate limits. Whether Lyria falls under Google Cloud's generative-AI indemnification list was not verifiable at the time of writing — confirm with Google for high-stakes commercial work.
Is AI music safe to use in commercial video in 2026?
The legal ground has shifted materially in twelve months. Universal Music Group settled its suit against Udio on October 29, 2025, in a deal reportedly combining a compensatory settlement, licensing agreements, and a joint licensed AI music platform slated for 2026. Warner Music reportedly settled with both Suno and Udio in late 2025, while Sony Music's suits against both companies and UMG's case against Suno were still in litigation as of mid-2026, with a pivotal fair-use ruling pending. The practical read: the market has split into licensed-training vendors (ElevenLabs, with its Merlin and Kobalt deals), provenance-first vendors (Google, with SynthID on every Lyria output plus artist-voice blocks), and vendors that grant commercial rights without disclosing training provenance. Google sits firmly in the provenance camp — it does not market Lyria as trained on licensed catalogs, but every output is watermarked and the guardrails are strict. Several of these cases remain unresolved, so treat this as the state of play as of August 2026, not settled law.
How do you prompt Lyria?
Lyria responds to concrete musical description: genre, mood, instrumentation, and tempo beat vague adjectives. "Upbeat indie-pop, female vocal, jangly guitars, handclaps, 120 BPM, sunny and nostalgic" outperforms "happy song." For Lyria 3, use the documented section tags and timestamps to shape structure, and write your prompt in the language you want the lyrics in, since lyrics follow the prompt language. Because generation is single-turn, prompting is your only editing pass: change one variable per regeneration rather than rewriting the whole prompt.
What can you make with Google Lyria?
The obvious pairing is music-driven video. Score a music video with a Lyria 3 Pro track built to your clip's length; because Lyria 3 generates timed lyrics, it also feeds naturally into a lyric video workflow where words need to land on beat. Shorter jobs — ad beds, podcast stingers, 30-second intros — are what Lyria 3 Clip targets, and lyria-002 covers instrumental beds where vocals would fight a voiceover.
What else people ask about Lyria
Is Google Lyria free? No. Official Gemini API pricing is $0.04 per song for Lyria 3 Clip and $0.08 for Lyria 3 Pro as of August 2026.
Can Lyria generate vocals and lyrics? Lyria 3 (Clip and Pro) and Lyria 3.5 generate vocals with timed lyrics, in the language of your prompt. lyria-002 is instrumental-focused.
How long can a Lyria song be? 30 seconds for lyria-002 and Lyria 3 Clip, up to 184 seconds for Lyria 3 Pro, and up to 3 minutes for Lyria 3.5 per DeepMind's model page.
Does Lyria watermark its music? Yes — every output carries an inaudible SynthID audio watermark, per Google's official documentation.
Can you edit a Lyria track after generating it? Not iteratively. Google's docs state music generation is single-turn; to change a track, you regenerate with a revised prompt.
Will Lyria imitate a specific artist? No. Prompts requesting a named artist's voice or copyrighted lyrics are blocked.
Where can you use Google Lyria?
You don't need a Google Cloud account to put Lyria in a video. The family runs natively in the Music tab of the invideo agent — pick Google Lyria 3 Pro, Lyria 3 Clip, or Lyria-002 from the model picker, generate the track, and cut it against your footage in the same session. The Lyria hub page tracks the current lineup, and the music video and lyric video workflows are the fastest routes from a Lyria track to a finished edit. Lyria is the Google-side music option in invideo, the AI video platform that gives serious creatives every major model in one place.
Version history: first published August 2026, covering lyria-002 (GA May 2025), the Lyria 3 preview pair (March 2026), and Lyria 3.5. Will be updated when Lyria 3 exits preview or Lyria 3.5 gets a published Vertex model ID.