Lyria 3.5 is reaching more creators through Gemini and its API. The eye-catching developer price is $0.08 for a full song. The more useful budget question is how many attempts you need before one track fits the edit, survives review, and is worth delivering.
Google’s September 4 announcement expands access through the Gemini app and Gemini API. The app offers genre and style choices, templates, vocals or instrumentals, and short or longer tracks. Google describes worldwide web and mobile availability and also names Flow Music, AI Studio, and Vids. This is an availability story, not a claim that the model was first invented or demonstrated today.
Choose the app or API for the job you actually have
| Route | Useful starting point | Check before committing |
|---|---|---|
| Gemini app | Exploring a brief and listening to candidate tracks | The features and limits shown in your account |
| Gemini API | Building a repeatable generation workflow | Billing, regional eligibility, model status, and error handling |
| Production edit | Selecting and fitting audio to a finished asset | Duration, transitions, mix, provenance, and client requirements |
I would start in the app when I need to discover the musical direction. I would move to an API workflow when I need a record of prompts, outputs, costs, and review decisions across repeated jobs. Do not assume that a consumer subscription includes developer usage, or that a globally announced app rollout means an unlimited API allowance.
The API specification is narrower than the Gemini feature list
The model page identifies lyria-3.5 as a preview. It accepts text and images and returns MP3 audio with lyrics text. Google lists 44.1 kHz stereo output. Caching, Batch API, Flex, Priority, and Live API are not supported for this model.
Those exclusions matter if you are reusing a generic Gemini integration. Do not carry over a batch-discount assumption or a realtime audio design without checking the model-specific table. Save the exact model identifier with each request so that a future change in the default does not silently alter your production record.
Calculate cost per accepted song, not per attempt
Google’s pricing table lists $0.08 per full song and no free API tier for Lyria 3.5. Here is the budget effect of rejecting some outputs. These are hypothetical selection rates, not measured model success rates.
| Illustrative run | Generation cost | Accepted tracks | Generation cost per accepted track |
|---|---|---|---|
| 10 generated songs | $0.80 | 5 | $0.16 |
| 10 generated songs | $0.80 | 2 | $0.40 |
| 10 generated songs | $0.80 | 1 | $0.80 |
The formula is useful even when the API bill is small. A client does not buy ten attempts; the client buys a usable asset. If listening, revisions, and fitting the track consume the afternoon, the generation fee is no longer the important line. Track hands-on minutes beside the accepted-song count, using the same logic as our AI video cost comparison.
Write a production brief, then define rejection rules
A request such as “make something cinematic” leaves too much room for an attractive but unusable result. My brief would state the role of the track, the intended duration, whether voices are allowed, the desired energy curve, and where the edit needs a quiet opening or a clean ending. These are acceptance criteria, not guaranteed controls.
Example brief: Create an original instrumental backing track for a product walkthrough. Keep the opening sparse enough for narration, build energy during the demonstration, and finish cleanly without an abrupt cutoff. Use light percussion and warm plucked textures. No lyrics, artist imitation, or prominent lead melody competing with speech. Aim for the length of the supplied edit brief.
Then reject a take for a concrete reason: it masks speech, introduces an unwanted vocal, changes mood at the wrong moment, or cannot end naturally in the available space. Save that reason. After several rejected takes, revise the brief instead of repeating the same request and hoping for a different answer.
A generated track still needs a review path
The music-generation guide describes single-turn generation rather than iterative editing of a previous output. It also describes nondeterministic results, SynthID watermarking, and restrictions involving artist voices and copyrighted lyrics. Do not design a client workflow around guaranteed identical reruns or precise edit-in-place behavior that the guide does not promise.
Keep the original output, prompt, model ID, generation date, and any later edits together. Listen to the full track, not just the first few seconds. Check the final exported video or audio because a track that works alone may fight with narration after compression or mixing.
Google’s API terms say it does not claim ownership over generated original content, while warning that similar content may be generated for others and leaving responsibility for use with the user. That is not a guarantee of exclusivity or blanket legal clearance. For a client job, assess the applicable terms and rights requirements. Our Firefly music rights and indemnity guide explains why those promises need to be read separately.
My verdict: test the selection workflow before automating volume
The access expansion makes it easier to try Lyria in a real creative process. I would first run one bounded brief, count the rejected takes, and measure how long it takes to place an accepted track into the final edit. Only then would I automate more requests. The winning workflow is the one that delivers usable audio with a clear record, not the one that creates the largest folder of MP3s.
What makes a track usable for your next project: mood, a precise ending, room for speech, or something else?
Checked September 4, 2026. Availability, specification, pricing, and policy descriptions come from Google. Cost examples are calculations, not benchmark results. Musthave.ai has not independently evaluated Lyria 3.5 audio quality.
