Lyria 3.5 is Google's flagship music generation model, a latent-diffusion model over temporal audio latents. It takes text and images as input and returns a full-length song — vocals or instrumental, with verses, choruses and bridges — as 44.1 kHz stereo MP3 plus the timed lyrics it sang. It reached the Gemini app and the Gemini API on September 4, 2026 under the model code lyria-3.5, in public preview.

How much does Lyria 3.5 cost?

$0.08 per song on the paid Gemini API tier, with no free tier as of September 2026. Pricing is per song, not per minute. For comparison, the legacy Lyria 3 Clip (30 seconds) is $0.04 and Lyria 3 Pro is also $0.08, while ElevenLabs' Eleven Music is $0.15 per minute — about $0.30 for a two-minute track.

How long can a Lyria 3.5 song be?

Google describes output as 'a couple of minutes' and publishes no maximum. There is no duration parameter — length is guided through the prompt (for example, 'create a 2-minute song') or via section timestamps, and Google says exact duration can be influenced rather than set. If you need exact-length audio, this is the wrong model.

Can I use Lyria 3.5 songs commercially?

Commercial use falls under the general Gemini API terms rather than a dedicated music licence, so review those terms for your jurisdiction and use case. Every generated track also carries a SynthID audio watermark, and prompts naming a specific artist's voice or requesting copyrighted lyrics are blocked by safety filters.

Quick Answer

What Is Lyria 3.5? Google's $0.08 Song API Explained

Published: September 7, 2026

The Short Answer

Lyria 3.5 is Google’s flagship music generation model, and as of September 4, 2026 any developer can call it through the Gemini API at $0.08 per full song.

The model is not new — Google launched it inside Flow Music on July 29, 2026. What changed on September 4 is access: it landed in the consumer Gemini app with genre templates and short-or-long track options, and in the Gemini API under the model code lyria-3.5, in public preview.

Underneath, it is a latent-diffusion model over temporal audio latents. It accepts text and images as input and returns an MP3 plus the lyrics it sang.

Full Specification

Property	Lyria 3.5
Model code	`lyria-3.5` (public preview)
Inputs	Text and images
Outputs	Audio (MP3, 44.1 kHz stereo) + text (lyrics)
Structure	Verses, choruses, bridges; vocals or instrumental; timed lyrics
Length	”A couple of minutes”; no published maximum, no duration parameter
Price	$0.08 per song, paid tier only — no free tier
Editing	Single-turn; no iterative editing of a generated track
Safety	Blocks prompts naming an artist’s voice or requesting copyrighted lyrics
Watermark	SynthID audio watermark on every track
Not supported	Function calling, structured outputs, caching, Live API, batch

That last row is the one people miss. No batch, no caching, no structured outputs, no Live API — Lyria 3.5 is a plain request/response generation endpoint, not a composable piece of a larger Gemini pipeline.

The Four Limits That Actually Bite

1. There is no duration parameter. You ask for a two-minute song in the prompt and get something around two minutes. Google’s own wording is that exact duration “can be influenced,” not set. For a podcast bed or a reel that has to land on a hard cut, that is a real workflow problem — you generate, measure, and regenerate.

2. Generation is single-turn. There is no “make the chorus louder” round trip. You cannot edit a generated track; you re-prompt and regenerate the whole thing. At $0.08 a shot that is cheap in money and expensive in iteration time.

3. Safety filters are strict by default. Prompts that name a specific artist’s voice or ask for copyrighted lyrics are blocked. This is the right default for a model called by millions, and it is a constraint you design around rather than argue with.

4. Every track is watermarked. SynthID is embedded in all output. That is a feature for provenance and disclosure compliance, and a consideration if your use case assumes unmarked audio.

Price in Context

Lyria 3.5 is priced per song, which is the first thing to notice when comparing it to per-minute models.

Model	Unit price	Two-minute track
Lyria 3.5 (Google)	$0.08 per song	$0.08
Lyria 3 Pro (Google, legacy)	$0.08 per song	$0.08
Lyria 3 Clip (Google)	$0.04 per 30s clip	$0.16 (4 clips, stitched)
Eleven Music (ElevenLabs)	$0.15 per minute	$0.30

At roughly a quarter of ElevenLabs’ price for a two-minute track, Lyria 3.5 looks like an easy win. It is not a like-for-like comparison. Eleven Music returns MP3 or WAV, supports section-level edits after generation, accepts a ~30-second audio reference to steer style, is multilingual, and its paid plans state explicit commercial clearance. Google’s model returns MP3 only, edits nothing after the fact, and leaves commercial terms to the general API terms.

A quarter of the price buys a different product, not a cheaper copy of the same one.

What Google Did Not Publish

No numeric benchmarks. The model card describes human and automated evaluation across music quality and aesthetics, vocal quality, audio fidelity, and prompt adherence — and reports the results qualitatively: gains over Lyria 2 on audio fidelity and on lyric prompt adherence. There is no published score, no head-to-head, no MOS figure.

If you are choosing between music models in September 2026, there is no vendor-published number to compare. You have to run your own listening test on your own prompts. Budget for that.

Who Should Use It

If you…	Verdict
Need cheap background scoring at volume	Yes — $0.08/song is the cheapest full-song API here
Need exact-length tracks	No — no duration parameter
Need WAV or stems	No — MP3 only
Need to revise a generated track	No — single-turn only
Need explicit commercial music licensing	Check — falls under general Gemini API terms
Need image-conditioned music	Yes — image input is supported, unusually

Image input is the underrated feature. Feeding a keyframe or a brand board in alongside the text prompt is a genuinely different steering mechanism from the text-plus-audio-reference approach ElevenLabs uses.

Last verified: September 7, 2026.