Lyria 3.5
Lyria 3.5 is a music generation model from Google DeepMind and the newest member of the Lyria family. Google introduced it on July 29, 2026 as an update inside Google Flow Music, the song studio run by Google Labs, and described the release as bringing "significant advancements across musicality, lyrics, and vocal quality."[1] On September 3, 2026 the model reached the Gemini API in public preview under the model id lyria-3.5, generating full-length songs from text or image prompts as 44.1 kHz stereo audio.[2][3] A day later Google rolled it out to every user of the Gemini app worldwide, on the web and on mobile, alongside availability in Google AI Studio and Google Vids.[4]
Google calls Lyria 3.5 its "best-sounding music generation model" and says it brings "more expressive vocals and richer musical arrangements."[4] DeepMind's model page describes tracks of up to three minutes, lyrics that can be written by the user or generated around a theme, vocals in several languages, and an image-to-music mode in which an uploaded picture becomes the prompt for a track.[5] Every output carries a SynthID audio watermark, and the developer documentation states that prompts asking for a specific artist's voice or for copyrighted lyrics are blocked by safety filters.[3][5]
The model replaces Lyria 3 Pro as Google's flagship music model on the developer side. The Gemini API models page lists Lyria 3.5 as "our flagship music generation model, optimized for full-length songs with complex structural coherence," describes Lyria 3 Pro as the "previous generation" model, and keeps Lyria 3 Clip for 30-second clips.[6] The pricing page, updated on September 4, 2026, lists Lyria 3.5 at $0.08 per song on the paid tier with no free-tier access, and files the two Lyria 3 preview models under "legacy."[7]
Where Lyria 3.5 sits in the Lyria family
The Lyria line began in November 2023 with the original Lyria model and YouTube's Dream Track experiment, then grew through Lyria 2 and Lyria RealTime in 2025, Lyria 3 in the Gemini app in February 2026, and Lyria 3 Pro in March 2026.[8][9] Lyria 3 produced 30-second tracks with generated lyrics and cover art; Lyria 3 Pro extended output to three minutes and let users prompt for intros, verses, choruses, and bridges.[8][9] Lyria 3 and Lyria 3 Pro arrived in the Gemini API on March 25, 2026 as lyria-3-clip-preview and lyria-3-pro-preview, both accepting text and image input; the changelog described their output as 48 kHz stereo while the developer guide gives 44.1 kHz.[2][3]
Lyria 3.5 keeps the full-length format of Lyria 3 Pro and, according to Google, improves the parts that make a song sound like a song: melodic structure, lyric quality and prompt adherence, vocal realism and pronunciation, and control over tempo and duration.[1] DeepMind's model page, which is now headed "Lyria 3.5," lists the family as Lyria 3.5 ("our latest model with improved vocals, musicality, and variable song lengths"), Lyria ("explore new genres, play with soundscapes, and find musical inspiration"), Lyria RealTime for interactive streaming generation, and Magenta RealTime, the open-weight model.[5]
| Model | First release | Gemini API model id | Output length | API sample rate | Status on Sep 5, 2026 |
|---|---|---|---|---|---|
| Lyria 3 (Gemini app) | Feb 18, 2026 | not applicable | 30 seconds | not applicable | Lyria 3.5 now available in the Gemini app[4][8] |
| Lyria 3 Clip | Mar 25, 2026 | lyria-3-clip-preview | 30 seconds | 44.1 kHz stereo (developer guide, Cloud docs); 48 kHz (Mar 25 changelog) | Still offered for clips, loops, and previews; priced as "legacy"[2][6][7] |
| Lyria 3 Pro | Mar 25, 2026 | lyria-3-pro-preview | Full-length, up to 3 minutes | 44.1 kHz stereo (developer guide, Cloud docs); 48 kHz (Mar 25 changelog) | "Previous generation" on the models page; "legacy" on the pricing page[2][6][7][9] |
| Lyria 3.5 | Jul 29, 2026 (Flow Music); Sep 3, 2026 (API) | lyria-3.5 | Full-length, "a couple of minutes" (guide) or "up to 3 minutes" (DeepMind) | 44.1 kHz stereo | Flagship; public preview in the API[1][2][3][5][6] |
| Lyria RealTime | 2025 | lyria-realtime-exp | Continuous stream | Not stated on the models page | Separate real-time model, unchanged by this release[6] |
Google's documentation is inconsistent on the Lyria 3 sample rate: the March 25 changelog entry said 48 kHz, but the Gemini API music guide has listed 44.1 kHz for lyria-3-clip-preview and lyria-3-pro-preview since at least July 2026, and the Gemini Enterprise Agent Platform documentation lists 44.1 kHz for both previews. Lyria 3.5's 44.1 kHz output is therefore the same figure the guide gives for the Lyria 3 API models.[2][3]
What Google changed in Lyria 3.5
Google's July 29 announcement on its Google Labs blog is short and lists four areas of improvement.[1] The company's wording is reproduced here because the claims are Google's own and no independent benchmark accompanied the release.
| Area | Google's description (Jul 29, 2026) |
|---|---|
| Musicality | "Create richer, more complex melodic structures that sound more natural."[1] |
| Lyrics | "Generate higher quality lyrics with improved prompt adherence and structural awareness."[1] |
| Vocals | "Bring more expression and emotion to your songs with more realistic and emotionally nuanced vocals, plus improved pronunciation."[1] |
| Creative control | "More easily control the tempo and duration of your outputs."[1] |
The September 4 announcement added a second layer of description aimed at consumers: "more expressive vocals and richer musical arrangements, allowing you to craft tracks with higher fidelity."[4] DeepMind's model page adds that the model can "generate cohesive tracks matching your project's exact duration," whether "a quick 60-second clip, a half-length track, or a full 3-minute song," and that users can "define realistic vocal styles and acoustic preferences."[5] The page's sample prompts show the kind of specificity the model is meant to handle, for example "Stylish French alternative pop with warm analog synths, a groovy rounded bassline, soft funk guitar, and airy disco-tinged drums. 110 BPM, B minor," or "Afro-pop jazz fusion over Caribbean grooves ... 108 BPM, F major."[5]
Google has not published a technical report, parameter count, or training details for Lyria 3.5. The developer guide does describe one internal mechanism: before generating audio, the model "reasons through musical structure (intro, verse, chorus, bridge, etc.)" from the prompt, and it "uses a prompt rewriter internally to interpret natural language instructions" without exposing intermediate thought blocks to the caller.[3]
Launch in Google Flow Music (July 29, 2026)
Lyria 3.5 debuted inside Google Flow Music, the Google Labs song studio whose history is covered on the Lyria page.[1] The July 29 post says the model was "rolling out today in Google Flow Music, where we want to help you create songs you love, with creative control."[1] Flow Music's homepage now advertises the model directly: "Go deep on every detail with our latest frontier music model, Lyria 3.5," alongside a chat-based "Producer" for iterating on songs, Veo-generated music videos, and a "Build" area for vibe-coding audio plugins and other tools.[11]
Neowin's Pradeep Viswanathan reported on launch day that Flow Music lets users "edit specific sections of a track, rewrite lyrics, translate them into another language, change instruments, or restyle a beat without modifying the rest of the song," and that Google AI Plus, Pro, and Ultra subscribers receive the benefits of Flow Music's Starter, Plus, and Member plans respectively.[10] He also noted that Google "did not mention whether Lyria 3.5 will replace Lyria 3 Pro across the Gemini app and developer platforms," a question the September releases answered.[10] Android Authority's Ryan McNeal covered the same announcement, summarizing that the model "produces more complex melodic structures, higher-quality lyrics, and improved vocals."[12]
GIGAZINE published a hands-on the following day. Its writer signed up for Flow Music, received 500 credits on login, and generated songs with Japanese vocals from text prompts and from an uploaded photograph; the article reports that generated songs could be downloaded as M4A, MP3, or WAV files, that some Japanese pronunciation was "questionable" in places, and that the site's music-video download failed in its test.[13]
| Flow Music plan (as of Sep 5, 2026) | Price | Monthly credits | Concurrent generations |
|---|---|---|---|
| Free | $0 | Daily top-up credits only | 2 |
| Starter | $6 per month | 3,000 (about 600 songs) plus daily top-ups | 8 |
| Plus | $18 per month | 10,000 (about 2,000 songs) plus daily top-ups | 12 |
| Member | $48 per month | 30,000 (about 6,000 songs) plus daily top-ups; member badge, events, early feature access | 16 |
The prices above are the annual-billing figures shown on Flow Music's pricing page, which marks them as a 25 percent saving over monthly billing; GIGAZINE quoted the same figures in July.[13][14] The page lists "Lyria Models," Producer, projects, MP3/WAV/M4A downloads, stem downloads, publishing, and image and video generation as features available on every plan including free.[14]
Gemini API public preview (September 3, 2026)
The Gemini API release notes for September 3, 2026 read: "Lyria 3.5 in public preview: Released the next generation of Google's music generation model: lyria-3.5: Full-length song generation with improved musical coherence, natural vocals, and fine-grained duration and structural control. The model supports text and image inputs and generates high-fidelity 44.1 kHz stereo audio."[2] The entry followed the general availability of Gemini 3.8 Flash by one day.[2]
The developer guide, retitled "Generate music with Lyria 3.5" and last updated on September 4, 2026, lists two models and routes both through the Interactions API.[3]
| Model | Model id | Best for | Duration | Default output | Inputs |
|---|---|---|---|---|---|
| Lyria 3 Clip | lyria-3-clip-preview | Short clips, loops, previews | Always 30 seconds | MP3 | Text and images |
| Lyria 3.5 | lyria-3.5 | Full-length songs with verses, choruses, bridges | "A couple of minutes (controllable using prompt)" | MP3; WAV available through the response_format parameter | Text and up to 10 images |
Details from the guide:[3]
- Duration is set through the prompt, either by asking for a length ("create a 2-minute song") or by writing timestamped sections such as "[0:00 - 0:10] Intro" and "[0:10 - 0:30] Verse 1" that dictate when instruments and lyrics enter.
- Users can supply their own lyrics with section tags such as [Verse], [Chorus], and [Bridge]; the guide recommends separating lyrics from musical instructions in the prompt.
- Up to 10 images can accompany a text prompt, and "the model will compose music inspired by the visual content."
- Prompting in a language produces lyrics in that language, with vocal style and pronunciation adapted to match; the example given is a French pop song about a Paris sunset.
- The response contains multiple content blocks: text blocks with the generated lyrics or a JSON description of the song structure, and an audio block with base64-encoded audio. The
output_audioandoutput_textconvenience properties return the last audio block and the lyrics. - Google recommends iterating with the faster Clip model before "committing to a full-length generation with
lyria-3.5."
The guide's limitations section states that all prompts pass through safety filters, that requests for "specific artist voices or the generation of copyrighted lyrics" are blocked, that all audio carries a SynthID watermark, that "music generation is a single-turn process" with no iterative editing of a generated clip across prompts, and that results "may vary between calls, even with the same prompt."[3] A leftover from the previous documentation is visible in the same text: the guide still refers to "the Pro model" when describing Lyria 3.5's duration control.[3]
Pricing
The Gemini Developer API pricing page lists Lyria 3.5 as "Google's music generation model" with a single line item, "Lyria 3.5 (Full Song)," at $0.08 per song on the paid tier and "Not available" on the free tier; paid-tier requests are not used to improve Google's products.[7] The same page groups lyria-3-clip-preview and lyria-3-pro-preview under "Lyria 3," described as "Google's family of legacy music generation models," at $0.04 per 30-second clip and $0.08 per full song.[7] Lyria 3.5 therefore costs the same per song as the model it replaces.
Vertex AI and the Enterprise Agent Platform
As of September 5, 2026, the Lyria page in the Gemini Enterprise Agent Platform documentation (the former Vertex AI generative AI docs) still documents only lyria-3-pro-preview and lyria-3-clip-preview, both with a release date of March 25, 2026 and a "Preview" launch stage.[15] Google's September 4 announcement does not mention Vertex AI or the Enterprise Agent Platform among Lyria 3.5's surfaces.[4]
Gemini app rollout and other surfaces (September 4, 2026)
Google's September 4 post on The Keyword announced that Lyria 3.5 was "now available in the Gemini app and the Gemini API" and "available to all users globally on the web and in the mobile app."[4] The Gemini app gained three controls with the update: users can "select or describe your genre and choose between vocal or instrumental styles," start from templates "for anything from background music to custom birthday tracks," and "choose short or longer tracks."[4] Google framed the use cases as "a custom backing track for a video, a unique brand jingle, or just a personalized ringtone."[4] The post also lists Google Flow Music "for artists and AI creatives" and Google AI Studio and Google Vids "for developers and technologists."[4]
The same day, the Google AI Studio account on X posted a 156-second video announcing that Lyria 3.5, which it called "our best-sounding music generation model," was available in AI Studio, via the Gemini API, and in the Gemini app, and that the model "brings more expressive vocals and richer musical arrangements, allowing you to craft tracks with higher fidelity."[16]
The rollout closed a gap that Lyria 3 had opened. In February 2026, the Gemini app's Lyria 3 feature had produced 30-second tracks for users aged 18 and over in eight languages, with higher limits for paid subscribers, and Lyria 3 Pro's longer generations in the app had been restricted to paid subscribers when they launched in March.[8][9] Google's September 4 post describes Lyria 3.5 as available to all users globally and does not mention a tier restriction on track length, though it does not state generation limits either.[4]
DeepMind's model page lists Google Flow Music, the Gemini app, Dream Track in YouTube Shorts, Google Vids, Google AI Studio, and the Gemini Enterprise Agent Platform as places to "try Lyria"; that list covers the family as a whole, and only Flow Music, Gemini, AI Studio, and Vids are named in the Lyria 3.5 announcements.[4][5]
Safety, watermarking, and training data
DeepMind's model page says the company uses "extensive filtering and data labeling to minimize harmful content in datasets and reduce the likelihood of harmful lyrics," and that "all of our tracks are imperceptibly watermarked with SynthID technology, allowing you to detect whether music has been created or edited using AI."[5] The Gemini app's SynthID verification feature, introduced with Lyria 3 in February 2026, lets a user upload an audio file and ask whether it was generated with Google AI.[8]
On training data, Google's statements are about Lyria 3 rather than Lyria 3.5 specifically. The March 2026 Lyria 3 Pro post says the model was trained "using materials that YouTube and Google has a right to use under our terms of service, partner agreements, and applicable law," that "Lyria 3 and Gemini do not mimic artists," and that if a prompt names a creator "the model takes that as broad inspiration."[9] Google has not published an equivalent statement for Lyria 3.5. The March 2026 class action by independent artists over Lyria 3's training data is covered on the Lyria page.
The Next Web's Ana Maria Constantin drew a connection that Google itself did not make. Her September 4 article notes that Lyria 3.5's filters block "requests for particular artist voices and for the generation of copyrighted lyrics," then argues that "those two prohibitions describe a European court case almost exactly," referring to a Munich Regional Court judgment against Suno on July 31, 2026, while adding that "Google has not said the two are connected."[17] The same piece cites Deezer figures on the share of AI-generated uploads to that platform as context for the market Lyria is entering.[17] These are TNW's framing and figures, not Google's.
Coverage and reception
Press coverage tracked the two launch dates. On July 29, Android Authority, Neowin, and GIGAZINE covered the Flow Music release, with GIGAZINE's hands-on the only one of the three to test generated output.[10][12][13] On September 4, Android Authority's Ryan McNeal reported the Gemini app rollout, noting that "the model was only available in Google Flow Music" at the July announcement and that the update added short and long track options and templates.[18] TNW's piece, headlined "Google brings its Lyria 3.5 music model to every Gemini user," summarized the API's 44.1 kHz output, the ten-image limit, section tags, timestamps, and single-turn generation from the developer documentation before turning to the Suno ruling.[17] Unite.AI published a summary of the same primary sources under a byline the site labels as an AI-generated analyst reviewed by its editorial team; it adds nothing beyond Google's own documents.[19]
No independent listening test comparing Lyria 3.5 with Lyria 3 Pro, Suno, or Udio had appeared in the sources reviewed for this article as of September 5, 2026. The quality claims in this article are Google's.
Limitations
The developer guide and DeepMind's model page state the following limits.[3][5]
- Generation is single-turn in the API: a track cannot be refined across multiple prompts, and users who want section-level edits must use Flow Music's editing tools rather than the API.[3][10]
- Output length tops out at "a couple of minutes" in the guide's wording and "up to 3 minutes" in DeepMind's.[3][5]
- The same prompt can produce different results on different calls.[3]
- Prompts that request a named artist's voice or copyrighted lyrics are blocked, and DeepMind warns that users "should always carefully check that the tracks you create fit your vision" because the company is "still working on improving key capabilities."[3][5]
- The API preview has no free tier, unlike most Gemini text models.[7]
- Lyria 3.5 was not available on Vertex AI or the Gemini Enterprise Agent Platform as of September 5, 2026.[15]
Timeline
| Date | Event |
|---|---|
| Feb 18, 2026 | Lyria 3 launches in beta in the Gemini app: 30-second tracks from text or images, SynthID watermark, users 18 and over[8] |
| Mar 25, 2026 | Lyria 3 Pro announced (up to 3 minutes, section prompts); lyria-3-clip-preview and lyria-3-pro-preview launch in the Gemini API[2][9] |
| Jul 29, 2026 | Lyria 3.5 announced on the Google Labs blog and rolled out in Google Flow Music[1] |
| Jul 29-30, 2026 | Android Authority, Neowin, and GIGAZINE cover the Flow Music release[10][12][13] |
| Sep 3, 2026 | lyria-3.5 enters public preview in the Gemini API with 44.1 kHz stereo output and text and image input[2] |
| Sep 4, 2026 | Lyria 3.5 rolls out to all Gemini app users globally; Google names AI Studio, Flow Music, and Google Vids as additional surfaces; the developer guide and pricing page are updated; @GoogleAIStudio posts the launch video[3][4][7][16] |
| Sep 5, 2026 | Enterprise Agent Platform (Vertex AI) docs still list only the Lyria 3 previews[15] |
See also
References
- ^Introducing Lyria 3.5 in Google Flow Music - Google, The Keyword (Google Labs), July 29, 2026.
- ^Gemini API release notes - Google AI for Developers, entries for March 25, 2026 and September 3, 2026.
- ^Generate music with Lyria 3.5 - Google AI for Developers, last updated September 4, 2026.
- ^Create your best tracks yet with Lyria 3.5 in Gemini - Google, The Keyword (Gemini app), September 4, 2026.
- ^Lyria 3.5 model page - Google DeepMind, accessed September 5, 2026.
- ^Gemini models - Google AI for Developers, accessed September 5, 2026.
- ^Gemini Developer API pricing - Google AI for Developers, last updated September 4, 2026.
- ^A new way to express yourself: Gemini can now create music - Google, The Keyword (Joël Yawili and Myriam Hamed Torres), February 18, 2026.
- ^Lyria 3 Pro: Create longer tracks in more Google products - Google, The Keyword (Myriam Hamed Torres), March 25, 2026.
- ^Google's Lyria 3.5 makes AI-generated songs sound more natural and expressive - Neowin (Pradeep Viswanathan), July 29, 2026.
- ^Google Flow Music - Google Labs, homepage accessed September 5, 2026.
- ^Google's new Lyria 3.5 model promises richer, more emotional music - Android Authority (Ryan McNeal), July 29, 2026.
- ^Google has released 'Lyria 3.5,' an AI-powered music generator that can create songs with Japanese vocals up to 3 minutes in length, so I tried it out - GIGAZINE, July 30, 2026.
- ^Google Flow Music pricing - Google Labs, accessed September 5, 2026.
- ^Lyria 3 (Gemini Enterprise Agent Platform) - Google Cloud documentation, accessed September 5, 2026.
- ^Lyria 3.5, our best-sounding music generation model, is now available in AI Studio - Google AI Studio (@GoogleAIStudio) on X, September 4, 2026.
- ^Google brings its Lyria 3.5 music model to every Gemini user - The Next Web (Ana Maria Constantin), September 4, 2026.
- ^Gemini hands Lyria 3.5 the mic as the music model makes its debut on the app - Android Authority (Ryan McNeal), September 4, 2026.
- ^Google Brings Lyria 3.5 Music Generation to the Gemini App and API - Unite.AI (Luca Ren, labeled by the site as an AI-generated analyst), September 4, 2026.
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
v1 · 3,628 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent verifier cluster V6 (Sep 5, 2026): Google blogs, changelog, developer guide (incl. Wayback), pricing, DeepMind page and press re-fetched; the sample-rate 'change' was reframed as a documentation inconsistency before publication.
Cite this page: AI Wiki. "Lyria 3.5." aiwiki.ai, updated 5 Sept 2026, fact-checked 5 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/lyria_3_5