Lyria 3 Clip Preview: 30-Second Music Generation at $0.04 a Clip

If you found Lyria 3 Clip in a model catalogue, the listing probably looked broken: a million-token context, a price of zero per token, and an output type of audio. None of that describes how the model is used or paid for. It is a music generator that returns one 30-second track per request, and Google bills it per clip.

Google launched it on 25 March 2026 alongside Lyria 3 Pro, both as public previews in the Gemini API. Clip is the short, fast one. This guide covers what that fixed length is good for, what it costs, and how to work with a model whose output you cannot read.

Billing is per clip, not per token

Google's pricing page lists Lyria 3 Clip Preview at $0.04 per song, paid tier only, with no free tier. Every request yields one 30-second clip, so the arithmetic is unusually simple: a thousand clips cost $40, regardless of how long or short your prompt is.

That is why a token-priced catalogue shows zeros. The per-token fields do not apply, and any aggregator listing that shows a non-zero token price for this model is worth checking against Google before you budget from it. The one cost variable you control is how many generations you throw away, and with music that number is higher than you expect, because the only way to judge a clip is to listen to it.

What 30 seconds is actually good for

Google's model page describes Clip as optimised for short musical clips, loops and previews, and the fixed length makes the fits obvious:

• Loops and beds — background audio for a game menu, an app onboarding screen or a product video, where the clip will be looped or faded anyway.
• Short-form video — 30 seconds covers most vertical-video formats without trimming.
• Prompt auditions — trying ten stylistic directions cheaply before committing to a full-length render on a more expensive model.
• Previews — letting a user hear a direction before they pay for, or wait for, a complete song.

What it is not for is anything with a structure longer than half a minute. You cannot ask Clip for two minutes; Google's documentation states the length is always 30 seconds. Stitching clips together rarely works either, because consecutive generations do not share key, tempo or arrangement reliably.

Calling it and reading the result

Lyria 3 accepts text and images as input. Google's music guide describes the model composing music inspired by the images you send, rather than describing them. Google's model page gives a 131,072-token input limit, well below what some catalogues list. The current Gemini API documentation calls the model through the Interactions API, client.interactions.create, and returns two things: base64-encoded audio and a text part.

interaction = client.interactions.create(
    model="lyria-3-clip-preview",
    input="Warm lo-fi hip hop loop, mellow Rhodes chords, vinyl crackle, 80 bpm. Instrumental only, no vocals.",
)
audio = interaction.output_audio   # base64 MP3
lyrics = interaction.output_text   # lyrics and structure

Output is stereo MP3. Google's pages disagree on the sample rate, 44.1 kHz in the music generation guide and 48 kHz on the model page, so check the files if it matters to your pipeline. The text part matters too: Google describes it as the generated lyrics or a JSON description of the song structure. Store it alongside the audio, because it is the only machine-readable description of what was generated.

Prompting a model you cannot read back

Text models fail visibly. Music models fail audibly, which means you need a listening step in the loop. A few habits reduce the number of wasted clips:

• State instrumental explicitly. This family is built around vocals and lyrics, and instrumental output is something you request. Google's own example adds "Instrumental only, no vocals" to get a bed.
• Name tempo and instrumentation, not adjectives. "90 bpm, upright bass, brushed drums" is more reproducible than "chill and jazzy". Google's example prompts on its Lyria page state tempo in bpm.
• Expect variation. Google's docs note that identical prompts can return different results. Generate three, pick one, and log the prompt with the winner.
• Do not reference artists. Requests for a specific artist's voice or copyrighted lyrics are blocked by safety filters.

Generation is single-turn. You cannot ask for the same clip with a slower tempo; you send a new prompt and get a new clip. Design your interface around regeneration rather than editing.

Watermarks and rights

Every track carries an inaudible SynthID watermark, which Google says lets the audio be identified as AI-generated. If you publish generated music, assume it can be identified as such. Whether that matters depends on the platform you publish to and its disclosure rules, so check those before building a feature that auto-posts.

Preview status, and what replaced it

Clip is a preview model, and Google has since moved on. Lyria 3.5 became generally available on 3 September 2026, and Google's deprecations page lists it as the recommended replacement for both Lyria 3 previews, with no shutdown date announced yet. A preview can change behaviour or disappear on short notice, so treat lyria-3-clip-preview as something to prototype on, not something to hard-code into a product. Model deprecation and migration covers how to keep that swap cheap.

Note one thing before migrating: Google's current docs describe Lyria 3.5 as the full-length model and still show the Clip preview for 30-second output. If fixed short clips are your use case, confirm what the GA lineup offers for them at the time you migrate.

A workable pipeline

1. Audition directions with Clip at $0.04 each. Ten directions, three takes each, is $1.20.
2. Have a human pick. Automated audio scoring is not yet good enough to replace an ear for taste.
3. Render the chosen direction at full length on Lyria 3 Pro or 3.5 if you need more than 30 seconds.
4. Log prompt, image input, returned lyrics text and the chosen file together, so you can reproduce the direction later.