Lyria 3.5 can generate full songs, expressive vocals, structured lyrics, and polished stereo audio — and Google is now putting it directly inside Gemini.
September 4, 2026 — Google just made AI music a lot more mainstream.
Its latest music-generation model, Lyria 3.5, is now available in the Gemini app, Gemini API, Google AI Studio, and Google Vids.
The model itself is not brand new. Google first introduced it in Flow Music on July 29.
What changed today is distribution.
Instead of asking users to visit a dedicated music-generation product, Google is putting Lyria inside the assistant they already use for writing, coding, research, images, and everyday questions.
That is the real story.
Lyria 3.5 is designed to generate more than short musical clips.
It can create songs with:
Google says Lyria 3.5 outputs 44.1 kHz stereo audio, with MP3 as the default and WAV available through the API.
It can also accept text and image inputs, including multiple reference images, allowing music to be generated around a visual mood or creative concept.
That makes the model useful for much more than “make me a funny birthday song.”
Think video soundtracks, ad music, game loops, creator content, jingles, branded audio, and demo tracks.
Google says Lyria is available to all Gemini users globally across web and mobile.
Inside Gemini, users can describe the music they want, choose vocals or instrumental, select a genre, start from templates, and specify shorter or longer generations.
That dramatically lowers the barrier to entry.
Suno and Udio helped teach people that AI can generate songs from text.
Google's advantage is different: it does not necessarily need people to visit a music-generation website.
It can simply put music generation in front of people who are already using Gemini.
That could matter more than any individual benchmark.
Google's examples also show that Lyria performs best when users think more like producers than casual prompters.
Instead of:
Make me a cool pop song.
Think:
Funk-pop at 95 BPM with electric guitar, warm bass, female vocals, and a big melodic chorus.
Or:
Afro-pop jazz fusion with steel pan, brass, syncopated percussion, and scat vocals.
The model responds to details such as tempo, instrumentation, vocal character, groove, genre, and arrangement.
That makes prompting feel less like asking for a song and more like giving creative direction to a virtual studio.
The Gemini API is where Lyria 3.5 becomes significantly more interesting.
Developers can provide custom lyrics using section labels such as:
[Verse]
[Chorus]
[Bridge]
They can also specify when musical events should happen.
For example:
[0:00 - 0:10] Intro: soft lo-fi beat and vinyl crackle
[0:10 - 0:30] Verse: Fender Rhodes and gentle vocals
[0:30 - 0:50] Chorus: full band and soaring synths
[0:50 - 1:00] Outro: solo piano and fade
That opens the door to applications where music needs to fit a precise creative structure rather than simply sound good.
Google Vids is an obvious example.
Instead of searching stock libraries for something that roughly fits a video, users could generate music specifically for the scene, pacing, and mood.
Google is offering a companion model called Lyria 3 Clip.
Its model ID is:
lyria-3-clip-preview
It is designed for 30-second generations such as loops, previews, stings, and short social assets.
The basic split is simple:
Lyria 3.5: full songs.
Lyria 3 Clip: short-form audio.
Google is also emphasizing safety controls.
Lyria is designed to block requests for copyrighted lyrics and direct attempts to recreate named artists' voices.
Generated audio also uses Google's SynthID watermarking technology.
That gives Google a somewhat different product posture from some AI music competitors currently facing copyright pressure.
But there is still an important gap.
Google has not published detailed training-data disclosures for Lyria 3.5, nor has it released comprehensive public listening benchmarks against Suno, Udio, or even Lyria 3 Pro.
So while the output guardrails are becoming clearer, the training-data story remains much less transparent.
That question is harder to answer.
Google has not published a serious head-to-head comparison.
Early independent reactions suggest that Lyria 3.5 has improved vocals, stronger melodies, and better song structure than previous versions.
But familiar weaknesses remain.
Pronunciation can still be inconsistent. Some languages appear stronger than others. And listeners continue to complain that many AI songs share a recognizable synthetic polish — technically impressive, but somehow similar.
That may be Lyria's biggest challenge.
The technology is already good enough to generate songs.
The next frontier is making those songs feel less like they were generated.
AI music is moving out of the novelty phase.
And Google's launch gives us a glimpse of what comes next.
Music generation becomes another background capability inside larger products.
Need background music for a presentation? Generate it.
Need a soundtrack for a video? Generate it.
Need a jingle for a local business? Generate it.
Need a loop for a game prototype? Generate it.
Need a birthday song? Ask Gemini.
At that point, the interesting question is no longer whether AI can make music.
It clearly can.
The question becomes what happens when music generation is simply another button inside the software people already use.
That is where Lyria 3.5 gets interesting.
The model is impressive. But distribution may be Google's real advantage.
