Lyria 3.5 arrived on AIxploria at 03:10 UTC on 6 September 2026, carrying a “new AI tool” badge, a 4.5-star rating and a line in the corner of the card reading “#2 in Music”. The model behind that card is Google DeepMind’s music generation system, and the card is unusually well researched: it quotes a per-song API price, names the model architecture, and gets the release timeline right. Most directory entries do not manage any of that.
Two of its claims do not survive a check against Google’s own documentation, and they are the two that matter most to anyone deciding whether to build on this model commercially. This is a documents-first read of that listing, and it sits alongside our earlier reads of the GPT-6 Astra card and the Gemini 3.8 Flash listing in the same directory. If you are tracking model releases across vendors, our AI models and tools hub keeps the wider picture.
The short version: the card is right about the machine and wrong about the music business. Every technical statement we could check held up. The rank badge contradicts the directory’s own category page, the training data claim that carries the card’s verdict appears nowhere in Google’s Lyria 3.5 documentation, and the promise of free access “for every account” runs into Google’s own help pages.
Table of contents
- What the Lyria 3.5 Card Actually Claims
- The Rank Badge That the Directory’s Own List Contradicts
- How Long Lyria 3.5 Took to Reach AIxploria
- Lyria 3.5 Claim by Claim Against Google’s Documentation
- What Lyria 3.5 Actually Costs to Run
- The Licensed Music Claim Google Never Made for Lyria 3.5
- Is Lyria 3.5 Really Free for Every Account?
- Lyria 3.5 Against Suno and Udio
- The Verified Lyria 3.5 Capability Register
- What This Means If You Are Choosing Lyria 3.5
- How to Read Any AIxploria Card
- References and Further Reading
What the Lyria 3.5 Card Actually Claims
The AIxploria entry sits at aixploria.com/en/lyria-3-5-google-deepmind/ as post 51194, published 03:10 and modified 03:51 UTC on 6 September 2026. It is tagged Latest AI and Music, marked “Verified Tool”, priced “Freemium”, and lists 34 alternatives.
The numbers behind the star rating
The visible rating is 4.5 out of 5. The page’s structured data gives the vote count that produces it: ratingValue 4.5 from a ratingCount of 4. Four people. The card also carries 69 upvotes. A launch-week rating built on four votes is enthusiasm, not evidence, and reading the count rather than the stars is the single cheapest sanity check on any directory listing.
For contrast, the predecessor card for Lyria 3 — live since 19 February 2026 — shows 4.6 from five votes and 133 upvotes. Seven months of exposure bought it one extra vote.
The editorial block
Unlike a bare listing, this card carries a written review: a summary, a pros-and-cons list, a prompt-behaviour table, a comparison against Suno, an access-surface list, four FAQ answers and a verdict. That verdict reads: “Three minutes of coherent song for eight cents, with training data on the right side of the courtroom.”
Both halves of that sentence are testable. One is correct. The other is the problem.
The Rank Badge That the Directory's Own List Contradicts
The card is badged “#2 in Music”, and that badge links to aixploria.com/en/category/music/. Following a badge’s own link is the cheapest finding available on any listing page, and here it pays immediately.
What sits at #2
The Music category page returns twelve server-rendered cards. Position two is held by Lyria 3 — the predecessor model, not the new one. Lyria 3.5 does not appear on that page at all, and it does not appear on page two either, which covers positions thirteen to twenty-four.
| Position | Tool on the Music category page | Note |
|---|---|---|
| 1 | Suno AI V5.5 | The card’s named rival |
| 2 | Lyria 3 | The slot the new card claims |
| 3 | Stable Audio 3.0 | Listed as an alternative |
| 4 | Eleven Music | Listed as an alternative |
| 5 | Google Flow Music | The surface Lyria 3.5 ships in |
| 6–12 | MiniMax Music 3.0, Udio, AIVA, Producer.ai, Mubert AI, Beatoven, AirMusic | Remainder of page one |
| 13–24 | YuE AI, Jammable, Vocal Remover, Similar Songs Finder, Rave DJ, Mureka, Covers AI, Songtell, AI Make Song, DiffRhythm, Song Generation Tencent, BeatViz AI | Page two — still no Lyria 3.5 |
Why this is a sharper failure than a missing entry
We found the same shape on the GPT-6 Astra card, badged “#1 in LLM models” while absent from that category. This one is worse in a specific way: the slot the badge claims is not empty. It is occupied by the model’s own predecessor, which is still described on its card as generating 30-second tracks.
A reader who trusts the badge, clicks through to compare, and finds Lyria 3 at number two has no way to tell whether they are looking at a stale list or a wrong badge. Both readings are available and the directory does not resolve either.
How Long Lyria 3.5 Took to Reach AIxploria
Our previous read of this directory measured how long a card takes to propagate into the ranked lists once it exists. This card lets us measure the step before that: how long the vendor’s release takes to become a card.
The upstream lag
Google announced Lyria 3.5 on 29 July 2026, in a post titled “Introducing Lyria 3.5 in Google Flow Music”. The DeepMind model card carries the same date. The AIxploria card appeared on 6 September — 39 days later, still badged “This is a new AI tool”.
That badge is defensible, though, because of a second date. Google opened the model to developers in AI Studio, the Gemini API and the Gemini app in early September; coverage of that developer launch is dated 4 September 2026. The card followed two days later. So AIxploria is tracking the developer availability, not the model release, and its own copy says exactly that — correctly.
The downstream lag, re-measured
We then re-ran the propagation greps that this series has been running since 3 September. Lyria 3.5 appears zero times on the Top 100 page and zero times on the Ultimate List. Lyria 3 appears five and eight times respectively.
That zero is expected rather than damning. Three days ago the same check returned zero for Claude Fable 5.1 and Gemini 3.8 Flash; today both return sixteen mentions on the Top 100 page and ten on the Ultimate List. A card takes roughly three days to reach the ranked views, and Lyria 3.5 is inside that window.
One methodological note worth repeating: the Free AI page returned nothing for any of these names, including Suno, so we excluded it. A page that returns zero for a tool you know is listed is client-rendered, and a zero from it means nothing.
Lyria 3.5 Claim by Claim Against Google's Documentation
We checked every testable statement on the card against four Google sources: the DeepMind model card, the launch post, the Lyria product page and the Gemini API documentation.
| Card claim | Google’s documentation | Verdict |
|---|---|---|
| Complete songs with verses, choruses, vocals and lyrics from a written prompt | Model card: outputs are “Audio (music), text (lyrics)” | Correct |
| Built on latent diffusion | Model card: “utilizes latent diffusion, applied to temporal audio latents” | Correct |
| $0.08 per song via the Gemini API | API price list: lyria-3.5 at “$0.08 per song” | Correct |
| Tracks up to 3 minutes, 44.1 kHz stereo | Product page: “up to 3 minutes long”; API docs: “44.1 kHz high-fidelity stereo” | Correct |
| No seed parameter, single-pass generation via the API | API docs: “Music generation is a single-turn process”; no seed documented | Correct |
| Prompt language sets the singing language | Product page: “Generate vocals in different languages” | Correct |
| Composes from uploaded images | Product page: “Upload an image and ask Lyria to transform it into a custom high-fidelity track” | Correct |
| Every output carries a SynthID watermark | Model card lists “applying SynthID watermarking” among mitigations | Correct |
| Released late July 2026, reached the API in early September | Launch post dated 29 July 2026; developer availability reported 4 September | Correct |
| Lyria 3 Clip charges $0.04 per 30-second clip | lyria-3-clip-preview, “$0.04 per song”, “up to 30 seconds” | Correct |
| Composes from “up to ten” images | Image input documented; no count stated on any Google Lyria page we checked | Unverified |
| “Tracks span 30 seconds to 3 minutes” | The 30-second floor is the fixed length of the separate Clip model; the product page’s short example is 60 seconds | Conflated |
| Pronunciation “clearly improved over Lyria 3” | Launch post says “improved pronunciation” with no baseline; the model card benchmarks against Lyria 2 | Unsourced baseline |
| “Trained on licensed music data” | The word “licensed” appears on none of the four Google Lyria pages | Not supported |
| Free “at no extra cost for every account” in Flow Music | Flow Music help page ties “our latest models” to a paid upgrade | Contradicted |
Ten of fifteen check out exactly. That is a better hit rate than the GPT-6 Astra card managed, and the technical half of this listing is genuinely reliable.
What Lyria 3.5 Actually Costs to Run
The card’s eight-cent figure is correct, and it is the most useful number on the page. What the card does not do is set it beside the rest of the model line.
| Model | API model ID | Price | Length |
|---|---|---|---|
| The new flagship | lyria-3.5 | $0.08 per song | Full-length, prompt-controlled |
| Lyria 3 Pro | lyria-3-pro-preview | $0.08 per song | Full-length, previous generation |
| Lyria 3 Clip | lyria-3-clip-preview | $0.04 per song | Up to 30 seconds |
| Lyria RealTime | lyria-realtime-exp | Not on the price list | Real-time streaming |
The detail the card omits
None of these models has a free API tier. The price list marks the free tier as unavailable for every Lyria entry, including the clip model. So the “Freemium” tag describes the consumer surfaces only, and a developer evaluating Lyria 3.5 pays from the first request.
The arithmetic that matters for a project
At $0.08 per song, a hundred generated takes cost $8 and a thousand cost $80. The clip model halves that to $0.04, which makes it the sensible way to audition a prompt before committing to a full render — exactly as the card suggests. Because there is no seed parameter, the same prompt does not reproduce the same track, so iteration count is a real budget line rather than a rounding error.
The Licensed Music Claim Google Never Made for Lyria 3.5
This is the card’s load-bearing claim. It appears twice — once in the pros list as “Trained on licensed music data”, and once in the body as “Google says its music models are trained on licensed data” — and it produces the verdict about training data being “on the right side of the courtroom”.
What the model card actually says
We searched all four Google Lyria surfaces for the string “licens”. It appears on none of them. The model card’s training-data section reads, in full: “Lyria 3.5 was trained on audio data. Audio datasets were annotated with text captions at different levels of detail.” The processing paragraph describes deduplication, safety filtering and quality filtering. It does not describe licensing.
Why the claim is not simply false
Google has made licensing statements about Lyria before, framed around YouTube. The company has said its music models were trained on material YouTube and Google had the rights to use under their terms of service, partner agreements and applicable law, and Lyria was developed in partnership with YouTube’s Music AI Incubator, with artists including Charlie Puth and John Legend licensing their voices.
So the card is repeating something Google has said about the Lyria family. What it is not doing is telling the reader that the claim rests on a contested legal theory — that YouTube’s terms of service licensed the uses — or that this theory is being tested in court by independent musicians, or that reporting has described Google as having trained on copyrighted major-label recordings first and approached rights holders afterwards.
Why the distinction is the whole point
“Trained on licensed music data” reads as a settled fact about provenance. What it actually is, is one party’s characterisation of its own contractual position, and machine learning provenance is precisely the question the courts have not finished answering. A directory card that presents it as a checkmark in a pros column is handing readers the strongest possible reading of an argument that Google’s own Lyria 3.5 documentation declines to make.
Is Lyria 3.5 Really Free for Every Account?
The card’s FAQ answers “Is Lyria 3.5 free?” with “Yes, inside Flow Music, where the model is included at no extra cost for every account.”
What Google’s help pages say
Google Flow Music’s own help page states that “You can use all of Google Flow Music’s core features at no charge.” It then adds: “To access more credits, generations, and our latest models, upgrade your Google Flow Music subscription plan or Google AI membership plan.”
Lyria 3.5 is the latest model. Google does not say explicitly that it is gated behind a plan, but it does tie latest-model access to an upgrade, and the card resolves that ambiguity in the most generous direction available without citing anything.
The credit structure underneath
Flow credits are the real constraint. Every account gets 50 Flow credits per day with or without a subscription. Google AI Pro adds 1,000 credits a month, Google AI Ultra at $100 adds 10,000, and Ultra at $200 adds 25,000. The membership tiers map onto Flow Music plans: AI Plus to Starter, AI Pro to Plus, AI Ultra to Member.
| Access route | What you get | Cost |
|---|---|---|
| Flow Music, no subscription | Core features; 50 Flow credits per day | Free, credit-limited |
| Google AI Pro | Flow Music Plus; +1,000 credits per month | Subscription |
| Google AI Ultra ($100) | Flow Music Member; +10,000 credits per month | Subscription |
| Google AI Ultra ($200) | Flow Music Member; +25,000 credits per month | Subscription |
| Gemini app | Tracks with cover art; 18+; limits apply | Included, limited |
| Gemini API | Programmatic generation | $0.08 per song, no free tier |
The honest summary is that Lyria 3.5 is free to try and metered to use, which is what “Freemium” should mean. Unlike the GPT-6 Astra card, whose Freemium tag sat next to an FAQ answering “Is it free? No”, this card’s tag and its FAQ agree with each other.
Lyria 3.5 Against Suno and Udio
The card’s comparison section is short and mostly fair: “Suno v5.5 keeps the edge on length, up to 8 minutes, and adds voice cloning. Lyria 3.5 plays a different card, traceability and fine structural control.”
Where the framing goes stale
The card supports that framing by noting that Suno and Udio “face lawsuits from major labels over their training data”. That was a complete description a year ago. It is not one now.
Warner Music Group settled with Suno in November 2025 and established a licensing relationship covering future model training, with Suno’s current models to be phased out once licensed replacements are ready. Universal Music Group settled with Udio in October 2025, with a per-generation royalty reported at $0.002 to $0.005 and a joint platform planned. Sony has not settled, and litigation continues in Boston, where the plaintiffs moved in May 2026 to expand the case from 560 to 61,026 recordings.
What that does to the comparison
The card’s implied advantage is that a buyer choosing Lyria 3.5 avoids legal exposure its rivals carry. The reality is messier in both directions: two of the three major labels have signed licensing deals with the very companies the card points at, while the provenance of Lyria’s own training data is the subject of a proposed class action.
| Dimension | Google’s new music model | Suno v5.5 |
|---|---|---|
| Maximum track length | 3 minutes | 8 minutes, per the card |
| Voice cloning | Not offered | Offered, per the card |
| Watermarking | SynthID on every output | Not claimed on the card |
| Label litigation | Proposed class action over training data | Warner settled; Sony and Universal continue |
| Developer access | Gemini API, $0.08 per song | Own subscription service |
| Editing after generation | Section replace and extend in Flow Music; none via API | Not covered by the card |
The Verified Lyria 3.5 Capability Register
Stripping out everything we could not confirm, here is what Google’s own documentation supports.
Architecture and training
Lyria 3.5 is a music generation system that synthesises high-quality audio from a text prompt, using latent diffusion applied to temporal audio latents. It was trained on TPUs using JAX and ML Pathways, on audio datasets annotated with text captions at varying levels of detail.
Evaluation
Google’s evaluation framework was built with music experts and tests genre, mood and instrument concepts in and out of distribution. The stated results are that Lyria 3.5 improved significantly over Lyria 2 on audio fidelity, and demonstrated better prompt adherence with lyrics. Note the baseline: the model card does not publish a comparison against Lyria 3.
Controls and outputs
The launch post claims four improvements: richer melodic structures, higher-quality lyrics with better structural awareness, more expressive vocals with improved pronunciation, and easier control of tempo and duration. Outputs are audio plus lyric text, up to three minutes, in 44.1 kHz stereo, MP3 by default with WAV available.
Safety
Mitigations listed on the model card include dataset filtering, conditional pre-training, supervised fine-tuning, reinforcement learning from human and critic feedback, and product-level safety filtering with SynthID watermarking. In the Gemini app, tracks are watermarked and the feature is limited to users aged 18 and over. Google’s verification tooling now covers audio, so a file can be checked for SynthID on upload.
What This Means If You Are Choosing Lyria 3.5
For a video editor, podcaster or app developer, the practical read is straightforward.
Where it fits well
Three minutes of coherent, structured, watermarked audio for eight cents is a genuinely strong offer for background beds, stings and scratch tracks. If your work already lives in Google’s tools, the model is reachable from Flow Music, the Gemini app, AI Studio and the API without adding another vendor. Teams building this into a product should treat it the way they would any other model dependency in their AI strategy — as a component with terms attached, not a feature.
Where to be careful
Do not treat the “licensed training data” line as a commercial clearance. The card’s own FAQ is more careful than its pros list: it says commercial use “depends on the current terms of the surface you use” and that the SynthID watermark “proves a track’s AI origin but does not settle rights on its own”. That is the right answer, and it contradicts the verdict printed twenty lines below it.
Read the terms attached to your access route, budget for iteration because there is no seed parameter, and remember that the model is capped at three minutes.
How to Read Any AIxploria Card
Four checks take about five minutes and turn a listing into evidence.
Read the vote count, not the stars
The rating is in the page’s structured data alongside a ratingCount. Four votes produced this card’s 4.5.
Follow the rank badge’s own link
If a card claims a position in a category, open that category. Here the claimed slot belongs to the predecessor model.
Check the vendor’s date, not the card’s
A card’s publication date tells you when the directory noticed, not when the vendor shipped. The gap here was 39 days against the model release and two days against the developer release.
Search the vendor’s documentation for the card’s strongest adjective
The word doing the most work on this card was “licensed”. Searching Google’s four Lyria pages for it took one command and returned nothing, which is the finding this article is built on.
References and Further Reading
AIxploria — Lyria 3.5 listing
AIxploria — Music category page
Google DeepMind — Lyria 3.5 Model Card
Google — Introducing Lyria 3.5 in Google Flow Music
Google DeepMind — Lyria model family
Gemini — AI music and song generation
Gemini API — Music generation documentation
Google Flow Help — AI credits and Flow Music subscription plans
Google Flow Help — Manage your Google Flow credits
Music Business Worldwide — Google and ethical AI music development
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.