Google DeepMind’s music model is on the picker, and it is the cheapest song here
Google published Lyria 3.5 on fal this morning and it is on the sound box now. Hand it words and it writes a finished song: the music, the arrangement, and a voice singing words it wrote itself. Hand it a picture as well and it scores that picture.
It costs 3.5 credits a song. MiniMax Music 2.6, the other model here that sings, costs 5.25. We bought two songs on it before writing this, one on words alone and one with a picture wired on, and fal charged $0.1000 for each.
Three models that make music, and what each one is for
A song on Lyria is one price whatever you ask for. There is no length control, no quality step and no count, so a one-minute track and a three-minute track are the same 3.5 credits. CassetteAI is the cheapest thing here and it makes a bed to run under a scene, not a song. MiniMax Music sings, and it is the one that sings in Chinese.
| Model | Credits | What you get |
|---|---|---|
| CassetteAI Music | 0.70 a minute | A bed to run under a scene. No singing |
| Lyria 3.5 | 3.50 a song | A whole song with singing, any length you ask for in the words |
| MiniMax Music 2.6 | 5.25 a song | A whole song with singing, and it sings Chinese |
The number sits on the Generate button before you press it, same as every other box.
Try generating with Lyria 3.5It sings in eight languages, and Chinese is not one of them
Google lists English, German, Spanish, French, Hindi, Japanese, Korean and Portuguese. Ask it for a sung Mandarin line and you will not get one. That is the whole reason MiniMax Music 2.6 stays on the picker at a higher price: it is the model for a Chinese vocal, and for that job the extra 1.75 credits is what the job costs.
Everything else Lyria does works in any language, because it is only the SINGING that is limited. An instrumental brief written in Chinese is fine.
There is no length slider. You ask for the length in the words
Every other model here has a length you set with a control. Lyria has none, and that is the model rather than something missing from our box. Google’s own instruction is to put the length in the prompt: write “a 2-minute track”, or mark the sections out, “[0:00-0:30] Intro: soft piano” and so on.
So the box sends it your words and nothing else. Our two paid songs came back at 2 minutes 53 seconds and 1 minute 1 second, from briefs that asked for different lengths, and both cost the same 3.5 credits. Asking for a longer song is free.
Wire a picture onto the box and it scores the picture
This is the part no other music model here can do. Drop a still onto a sound box running Lyria and it writes music that matches the mood and the subject of that still. It is a reference, not a frame: nothing about the picture ends up in the file, because the file is a song.
One picture at a time, which is what Google publishes. The box shows it as a reference chip, the same chip a picture wired into a shot uses, and the price on the button does not move.
What comes back, and what is stamped into it
A 44.1 kHz stereo MP3, and the words it sang, which the box keeps alongside the track. Every file carries SynthID, Google’s own inaudible provenance watermark, which says a machine made it. That is Google’s stamp and it is in every Lyria file everywhere, not something we add.
Google’s safety filters refuse anything that would imitate a real artist’s voice, so asking for a named singer gets you a refusal rather than a track.
Free to start
40 credits when you sign up
40 free credits, yours for 14 days. Go make something great.
- SOTA models
- Infinite canvas
- Director stage
- MCP access
Got questions left?
Can it sing in Chinese?
No. Google lists eight languages and Chinese is not among them. For a Chinese vocal use MiniMax Music 2.6, which is 5.25 credits a song.
How do I make a longer song?
Ask for it in the words: “a 2-minute track”, or write the sections out with timestamps. There is no length control because the model does not have one, and a longer song costs the same 3.5 credits.
Does a picture cost extra?
No. One picture, and the number on the Generate button is the same 3.5 either way. We checked it against the bill: both songs we bought cost $0.1000, the one with a picture and the one without.
