# Lyria 3.5 Is in the Gemini API: $0.08 a Song, 44.1 kHz, and What It Means for AI-Made Video

> Google's Lyria 3.5 hit the Gemini app and API on September 4, 2026: full songs with vocals for $0.08, 44.1 kHz stereo, SynthID. Specs, limits, and how it compares.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-09-05
- Canonical (HTML): https://cellcog.ai/blog/lyria-3-5/
- Section: Guides / Choosing a platform
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- Google put Lyria 3.5 into the Gemini app and the Gemini API on September 4, 2026 (the model itself launched July 29 inside Google Flow Music). API status: public preview, model code lyria-3.5.
- Output is a full-length song with vocals or instrumental, verses, choruses and bridges, timed lyrics, 44.1 kHz stereo MP3. Length is 'a couple of minutes', steered by the prompt.
- Price: $0.08 per song on the paid API tier, no free tier. The older Lyria 3 Clip (30 seconds) is $0.04 and Lyria 3 Pro is also $0.08.
- Google publishes no numeric benchmarks. The model card reports qualitative gains over Lyria 2 on audio fidelity and on lyric prompt adherence.
- Limits from Google's own docs: no duration parameter and no published maximum (length is guided through the prompt, 'a couple of minutes'), single-turn generation (no iterative edits), safety filters that block artist-voice and copyrighted-lyric prompts, and a SynthID watermark on every track.
- For a two-minute track, $0.08 per song is roughly a quarter of ElevenLabs' $0.15 per minute; ElevenLabs answers with WAV output, section-level edits, audio references and explicit commercial licensing.
- For CellCog users nothing changes today: music generation keeps running on the model we ship, and Lyria 3.5 goes on the same test bench every music release goes on. The criteria are below.

## At a glance

- **What is Lyria 3.5?** Google DeepMind's music-generation model: full songs with vocals or instrumental arrangements, timed lyrics and song structure, output as 44.1 kHz stereo MP3. Latent diffusion over audio latents, per the model card.
- **What happened on September 4?** It reached the Gemini app (globally, web and mobile, with genre templates and short or longer tracks) and the Gemini API in public preview. The model had launched in Google Flow Music on July 29, 2026.
- **What does it cost?** $0.08 per full song on the Gemini API paid tier. There is no free tier. Google has not published a separate consumer quota for the Gemini app.
- **How good is it?** Google reports qualitative improvements over Lyria 2 in audio fidelity and lyric adherence, with no numbers. Judge it on your own prompts; below is the bench we use.
- **Does CellCog use it?** Not today. CellCog's music generation runs on the model we ship now, and every new music model is tested against it on the criteria in this post. If Lyria 3.5 wins on our bench, that change gets its own product update.

**Google's Lyria 3.5 is now in the Gemini app and the Gemini API, at $0.08 per full song.** The [announcement](https://blog.google/innovation-and-ai/products/gemini-app/better-tracks-lyria-gemini/) landed on September 4, 2026: Google's "best-sounding music generation model" gets genre templates and short-or-long tracks in the consumer app, and a `lyria-3.5` model code in the API, in public preview. The model itself is not new; Google [launched it in Flow Music on July 29](https://blog.google/innovation-and-ai/models-and-research/google-labs/lyria-3-5/). What is new is that any developer can call it.

That matters to us because music is load-bearing at CellCog. Every film, podcast and reel our AI employees produce has a score under it, generated from a prompt and mixed under the voice, and the [music model we ship](https://cellcog.ai/blog/text-to-music-generation/) is something our founder re-evaluates against every new release. So this post does two things: it puts everything Google actually published in one place, with links, and it says plainly what Lyria 3.5 would have to do to get into our pipeline.

## What shipped, from Google's own pages

Lyria 3.5 is a latent-diffusion model over temporal audio latents (Google's [model card](https://deepmind.google/models/model-cards/lyria-3-5/)). It takes text and images as input and returns an MP3 plus the lyrics it sang. The [API documentation](https://ai.google.dev/gemini-api/docs/models/lyria-3.5) and the [music generation guide](https://ai.google.dev/gemini-api/docs/music-generation) give the specifics.

*Table: Lyria 3.5 as documented by Google, September 4, 2026*

| Property | Lyria 3.5 |
|---|---|
| Model code | `lyria-3.5` (public preview) |
| Inputs | Text and images |
| Outputs | Audio (MP3, 44.1 kHz stereo) and text (lyrics) |
| Length | "A couple of minutes"; no maximum published, no duration parameter. Length is guided through the prompt ("create a 2-minute song") or section timestamps; Google says exact duration "can be influenced", not set |
| Structure | Verses, choruses, bridges; vocals or instrumental; timed lyrics |
| Price | $0.08 per song, paid tier only; no free tier |
| Editing | Single-turn: no iterative editing of a generated clip |
| Safety | Prompts naming an artist's voice or copyrighted lyrics are blocked |
| Watermark | SynthID audio watermark on every track |
| Not supported | Function calling, structured outputs, caching, Live API, batch |

Google is unusually candid about the two limits that matter most in production. Music generation "is a single-turn process," so there is no "make the chorus louder" round-trip; you re-prompt and regenerate. And every prompt runs through safety filters that block "specific artist voices or the generation of copyrighted lyrics." Both are the right defaults for a model that will be called by millions; both are constraints a pipeline has to design around.

## The pricing, next to the alternatives

The [pricing page](https://ai.google.dev/gemini-api/docs/pricing) lists Lyria 3.5 at $0.08 per song, with the previous generation still listed as legacy: Lyria 3 Clip (30 seconds) at $0.04 and Lyria 3 Pro (full song) at $0.08. Per song, not per minute, which is the first thing to notice when you compare it with a per-minute model.

*Table: What a two-minute track costs at list price, September 4, 2026*

| Model | Unit price | Two-minute track | Length and duration control |
|---|---|---|---|
| Lyria 3.5 (Google) | $0.08 per song | $0.08 | "A couple of minutes", prompt-guided; no duration parameter, no published maximum |
| Lyria 3 Pro (Google, legacy) | $0.08 per song | $0.08 | Same as 3.5 |
| Lyria 3 Clip (Google) | $0.04 per clip | 4 clips, $0.16, stitched | Fixed 30 seconds |
| Eleven Music (ElevenLabs) | $0.15 per minute | $0.30 | 3 seconds to 10 minutes, set exactly in milliseconds (`music_length_ms`); per-section durations via a composition plan |

The comparison is not one-dimensional. [Eleven Music](https://elevenlabs.io/docs/capabilities/music), at [$0.15 per minute](https://elevenlabs.io/pricing/api), returns MP3 or WAV, lets you edit the sound and lyrics of individual sections after the fact, accepts a ~30-second audio reference to steer style, is multilingual, and is "cleared for nearly all commercial uses" on its paid plans, per ElevenLabs' own page. Google's model returns MP3 only, edits nothing after generation, and leaves commercial terms to the general API terms. A quarter of the price buys you a different product, not a cheaper copy of the same one.

## What Google did not publish

No numbers. The model card describes human and automated evaluations across "music quality and aesthetics, vocal quality, audio fidelity, and prompt adherence," and reports two qualitative results: Lyria 3.5 "improved significantly compared to Lyria 2 on audio fidelity," and with lyrics it "demonstrates better prompt adherence" on simple and complex instructions. No MOS, no preference rate, no leaderboard. There is also no published consumer quota for the Gemini app, no stated commercial-use license on the model page, and no duration ceiling beyond "a couple of minutes."

None of that is disqualifying. It means the evaluation is yours to run, on your prompts, against the model you use today.

## What it means for CellCog users

Nothing changes today. CellCog's music generation keeps running on the model we ship, in both places it lives: a standalone tool that turns a prompt into a track of any exact length from three seconds to ten minutes, with a composition-plan mode that pins each section to an exact duration when a cut has to land on a beat, and the video pipeline that generates the score alongside the footage and [ducks it under every spoken line automatically](https://cellcog.ai/blog/seedance-2-5-full-films-one-prompt/). Our [podcast](https://cellcog.ai/blog/clocked-in-ai-employee-podcast/) runs on the same stack.

Lyria 3.5 goes on the bench every music model goes on. Five things decide it, in order:

- **Exact duration.** A video score has to be 47 seconds when the cut is 47 seconds. Google says length is steerable "using the prompt" and publishes no maximum; [ElevenLabs' API](https://elevenlabs.io/docs/api-reference/music/compose) takes an exact `music_length_ms` from 3,000 to 600,000, and per-section durations in a composition plan. Prompt-steered length has to hold within a second, repeatably, or it fails here first.
- **Instrumental-only reliability.** Most scores under narration must carry no vocals. A model that occasionally sings over the presenter is a model that costs a re-render.
- **How it sits under a voice.** Dynamic range and low-end density decide how much ducking a track needs. This is heard, not read off a spec sheet.
- **Vocal quality on lyric prompts.** Google's stated strength; also where the safety filter's artist-voice rule bites for anyone prompting "in the style of."
- **Licensing for our users' commercial work.** Our users ship what they make: ads, explainers, launch films. ElevenLabs states its commercial clearance on the product page; Google's terms have to be read for the same answer.

If Lyria 3.5 wins, the change ships as a product update with the receipts, the way our [model routing changes always do](https://cellcog.ai/blog/cellcog-smart-routing-gemini-3-8-flash/). If it loses, nothing happens and this post stays as the record of why we looked.

## The record

- **July 29, 2026**: Lyria 3.5 launches inside Google Flow Music ([Google Labs](https://blog.google/innovation-and-ai/models-and-research/google-labs/lyria-3-5/)). Model card published the same day.
- **September 4, 2026**: Lyria 3.5 reaches the Gemini app globally and the Gemini API in public preview ([Google](https://blog.google/innovation-and-ai/products/gemini-app/better-tracks-lyria-gemini/)); the [API changelog](https://ai.google.dev/gemini-api/docs/changelog) lists `lyria-3.5` with "fine-grained duration and structural control"; pricing page shows $0.08 per song, Lyria 3 models marked legacy.
- **Open**: numeric evaluations, a stated commercial-use license on the model page, a consumer quota for the Gemini app. We update this page if Google publishes any of them.

## FAQ

**Lyria 3 vs Lyria 3.5: what is the difference?**

Lyria 3 shipped as two API models: Lyria 3 Clip (30-second clips, $0.04) and Lyria 3 Pro (full songs, $0.08). Lyria 3.5 replaces the Pro tier as the current full-song model at the same $0.08, and Google now lists the Lyria 3 models as legacy. Google says 3.5 brings more expressive vocals and richer arrangements; it publishes no side-by-side numbers.

**Can Lyria 3.5 generate a song with my own lyrics?**

Yes. Google's guide says to separate the lyrics clearly from the musical direction in the prompt. Prompts that ask for a specific artist's voice or for copyrighted lyrics are blocked by the safety filters.

**How long can a Lyria 3.5 track be?**

Google publishes no maximum. The docs say 'a couple of minutes', guided through the prompt ('create a 2-minute song') or timestamps that define the structure, and that exact duration 'can be influenced', not set. There is no length parameter. The 30-second format belongs to the separate Lyria 3 Clip model. For comparison, ElevenLabs' music API takes an exact length from 3 seconds to 10 minutes.

**Is Lyria 3.5 music free to use commercially?**

Google's docs do not spell out a commercial-use license on the model page; usage is governed by the Gemini API terms, and every track carries a SynthID audio watermark. Check the terms for your use before shipping generated music in paid work.

**What is the cheapest way to get a two-minute AI music track in September 2026?**

On list prices, Lyria 3.5 at $0.08 per song is the lowest per-track price among the major APIs we track; ElevenLabs' Eleven Music is $0.15 per minute, about $0.30 for two minutes. Price is one axis: WAV output, editability, licensing and how the track behaves under a voice track decide the rest.

**Will CellCog switch its music generation to Lyria 3.5?**

Only if it wins on our bench, which tests exact-duration control for video sync, vocal quality, instrumental-only reliability, how the mix sits under speech with automatic ducking, and licensing for our users' commercial work. Our founder runs that evaluation on every music release, and a change would ship as a product update, never quietly.

## Related

- [CellCog Goes Musical: Text-to-Music Generation Is Here](https://cellcog.ai/blog/text-to-music-generation/index.md)
- [Full Films From One Prompt: Seedance 2.5 Powers CellCog Video](https://cellcog.ai/blog/seedance-2-5-full-films-one-prompt/index.md)
- [MiniMax H3 Max Renders Video Faster Than It Plays. That Is How AI Employees Get a Face.](https://cellcog.ai/blog/minimax-h3-max/index.md)

---

Markdown alternate of https://cellcog.ai/blog/lyria-3-5/. Try CellCog free, no credit card needed: https://cellcog.ai/signup