Lyria 3 Pro Preview: Full-Length AI Songs and What to Build On

Most generated music falls apart past the first chorus. The verse wanders, the hook never returns, and two minutes in the track sounds like a different song. Lyria 3 Pro is Google's attempt at the harder problem: a complete song, with verses and choruses that hold together, from a single request.

Google released it on 25 March 2026 as lyria-3-pro-preview, a public preview in the Gemini API, alongside the 30-second Lyria 3 Clip. Six months on, the more important fact is that Lyria 3.5 reached general availability on 3 September 2026 and Google names it as the replacement. That shapes how you should use Pro today.

What Pro adds over Clip

The difference is length and structure. Google's model page describes Pro as built for full-length songs, against Clip's fixed 30 seconds. It states no maximum length; Google's Lyria page describes tracks of up to three minutes, so test the lengths you need. Google's model pages and music guide describe a shared feature set:

• Vocals, with generated lyrics that follow the language of the prompt.
• Tempo, instrumentation and mood described in plain language.
• A text part returned with the audio, carrying lyrics or a description of the structure.
• Image input that steers the music.
• Structural coherence across verses, choruses and bridges — the claim that matters most for Pro.

Price doubles. Google's pricing page lists Lyria 3 Pro Preview at $0.08 per song against $0.04 for a Clip, paid tier only. Per second of audio, a song of two or three minutes is far cheaper than the same length in Clips, and unlike a run of Clips it is one coherent piece. If you need anything longer than 30 seconds, Pro is the only option in the Lyria 3 previews; if you need exactly 30 seconds, Clip is half the price.

Controlling structure

A full song is where prompt structure pays off. Google's music generation docs describe several controls. They demonstrate most of them with Lyria 3.5 rather than the Pro preview, so verify each one against Pro before you depend on it:

• Section tags such as [Verse], [Chorus] and [Bridge] to lay out the form.
• Timestamped sections such as [0:00 - 0:15] Intro: solo piano to control when instruments enter.
• Custom lyrics written into the prompt. Generated lyrics follow the language of the prompt.
• Duration requests in plain language, such as asking for a two-minute song.
• Instrumental output by asking for no vocals.

The response carries base64 audio plus a text part with lyrics and structure. Keep that text: it is the only searchable record of what a track contains, and it makes a generated library browsable later.

What it refuses, and what it marks

Three constraints shape any product built on it.

No imitation. Google's safety filters block prompts asking for a specific artist's voice or copyrighted lyrics. If your feature idea is "sounds like a named band", it will not work, and users will try it anyway — handle the refusal gracefully.

No editing. Generation is single-turn. You cannot ask Pro to keep the verse and rewrite the chorus. A change means a new generation that may differ in key, tempo and arrangement, so a "regenerate" button is the honest interface, not "edit".

Always watermarked. Every track carries an inaudible SynthID watermark, which Google says is there for identification. Generated songs can be identified as AI-made, which matters on any platform with disclosure rules.

Should you build on the preview now?

Probably not as the long-term target. Google's deprecations page lists lyria-3.5 as the recommended replacement for lyria-3-pro-preview. No shutdown date has been announced, but preview models carry no stability promise, and the dated GA model is what you can plan around.

The economics do not argue against moving. Google's pricing page lists Lyria 3.5 full songs at the same $0.08 per song. Google's current docs also describe 3.5 returning MP3 by default with WAV on request, at 44.1 kHz stereo, and accepting up to ten images alongside a prompt.

There is one case for staying a little longer: you have a library of prompts tuned on Pro and a release that cannot slip. Then pin the preview, schedule the migration, and rerun your prompt library on 3.5 before switching, because a newer music model will not interpret the same prompt the same way. Pinning model versions and model deprecation and migration cover the mechanics.

Costing a music feature

Per-song pricing makes forecasting easy, but the multiplier is regeneration, not usage. If users keep one song in four, your effective cost is $0.32 per kept song, not $0.08. Measure the keep rate in a beta before setting a price for users.

A cheaper pattern is to audition with Clip and render with Pro: generate three 30-second takes at $0.04 each, let the user pick a direction, then spend $0.08 on the full render. That is $0.20 per finished song with a choice built in, against $0.24 for three full Pro renders.

Evaluating output you have to listen to

There is no automated score worth trusting for musical quality, so build a small listening protocol:

1. Write fifteen prompts that match real user requests, including a few with custom lyrics and section tags.
2. Generate three takes of each. Variation between takes is part of what you are measuring.
3. Score each blind on three questions: does the chorus return, are the lyrics intelligible, does the ending land or simply stop.
4. Repeat on Lyria 3.5 with the same prompts. Blind model comparison covers keeping listeners from knowing which model they hear.