What happened
Suno's post announces three models: v6 and the experiment-minded v6-wild on the paid Pro and Premier tiers, plus v6-mini, which everyone gets. The text says: "As v6 rolls out, we will retire our previous models and move Suno entirely onto the v6 generation." So the older models go away, and no date is given.
The partner list, word for word: "we're introducing v6, a new generation of music models developed with our industry partners, including Warner Music Group, BMG and Believe." One of those partnerships is a day older: on 8 September Believe and TuneCore announced that "Believe will be part of the new music models we are developing with the music industry." They added that tracks made this way will be distributable on their platforms — the same two services that had earlier declared they would not distribute music made with the older models.
What the announcement says — and what it does not
That gap is a story in itself. Suno's official blog post never once uses the word "licensed", and it nowhere claims that v6 was trained on licensed data. All it says is that the model was developed "with our industry partners".
The "trained on licensed data" framing comes from the company's statements to the press: on that account v6 was trained from scratch, on different data than the earlier models, and that data is put together from material licensed from the partners, from user data and from in-house work — and it does not include Universal Music or Sony Music material. There will be revenue sharing as well, but no figure for it has been made public.
Two of the majors have no deal: Universal and Sony Music. And those two are precisely the ones suing the company.
There is no verifiable information about the training material: no catalogue, no volume, no audit. An artist opt-in right is in all three agreements, but its mechanism is not publicly documented, and by Suno's own wording the product built on it is still in development.
What you actually get if you use it in business
This is the point to read the terms of service rather than the announcement. Under the text in force since 3 September:
free no download, no commercial use
Pro 20 downloads/mo, commercial ok
Premier 60 downloads/mo, commercial ok
all tiers NO legal warranty
remix never for commercial use
The details that matter:
- Commercial use is tied to the download. The terms put it plainly: "You may not commercially exploit Output that has not been downloaded by you through an approved channel." Audio grabbed off the screen or obtained some other way does not count.
- The transfer of rights is conditional. Suno assigns "all right, title and interest" to the paying subscriber — and then adds straight away: "Due to the nature of machine learning, Suno makes no representation or warranty to you that any copyright will vest in any Output." You get what was theirs, in other words, with no guarantee that it is a copyright-protected work at all.
- There is no exclusivity. The terms say it outright: "Output may not be unique across users and the Service may generate the same or similar output for a third party."
- There is no legal warranty. The text expressly excludes any warranty that using the output will not infringe third-party rights — and beyond that, the user takes on an indemnity obligation, while the provider's liability is capped at the last six months' fees, or 100 dollars.
- Remix is a trap of its own. Remixes count as joint works and, whatever your subscription tier, may be used only non-commercially.
- What you acquired does outlive the subscription. Under the terms, what you downloaded and got rights to stays yours for good, even if you cancel later.
This is not legal advice, and it is not a peculiarity of one provider either: most generative services are built the same way. The substance is that at the contract level the legal risk is on the user, whatever data the model was trained on.
Where the lawsuits stand
| Case | Where it stands on 9 September 2026 |
|---|---|
| Universal + Sony v. Suno (Massachusetts, since June 2024) | Running. On 18 August the judge allowed the new claim built on "stream ripping", but refused to pull in a further 61,026 recordings — the case still covers the original 560 recordings. In its 1 September filing Suno admitted that it had taken audio from YouTube for training, but calls this "typical fair use". Fact discovery closes on 30 September, and dispositive motions are due in April 2027. |
| GEMA v. Suno (Munich, judgment of 31 July 2026) | The German collecting society won on six musical works. The court held that the training material had been absorbed into the model's parameters, and rejected reliance on the text and data mining exception. Under the judgment the provider is liable for the outputs, not the user. It is not final. |
| Isbell and others v. Suno (Massachusetts, 1 September 2026) | A new class action from four musicians — not on copyright grounds but on name and likeness rights, invoking seventeen state laws. |
The picture at the competitor is much the same: Udio has settled with Universal, with Warner, with Merlin and with the American music publishers, while in July Sony brought a fresh suit against it over more than thirty thousand recordings.
The pattern repeats: lawsuit first, deal second, new model generation in the meantime. This is how the market currently runs, not an exception to it.
The EU part
Two rules bear directly on this news.
Article 53 of the AI Act requires providers of general-purpose models to publish a "sufficiently detailed summary about the content used for training", following the AI Office template, and to have a copyright compliance policy that respects rights reservations. Models placed on the market before 2 August 2025 have a grace period until August 2027 — a model released in September 2026 does not fall under that grace period. We found no public training-data summary from Suno.
The rights reservation rests on Article 4 of the DSM Directive: the text and data mining exception holds only where the use has not been "expressly reserved by their rightholders in an appropriate manner, such as machine-readable means" in the case of content made publicly available online. Hungarian copyright law takes this over word for word.
On the collecting society side, the Spanish SGAE formally reserved the text and data mining rights over its repertoire in July 2024, and stated that reproducing the repertoire for AI development purposes requires a licence. Artisjus has published no such formal, machine-readable reservation; from January 2025, though, a work registration has to declare whether AI was used, and AI cannot be an author. Collecting societies in both countries forecast serious revenue losses through 2028.
What to do now
- If you use AI music in business, check which tier you are on. The free tier carries no commercial rights and no download either — so whatever you take from there and run under an ad is not lawful even by the provider's own terms.
- Keep the proof of the download. Commercial rights are tied to the download. An invoice, a download log and the file's metadata together answer the question "where did you get this".
- Count on the model being swapped out. The older models are being retired. If a brand sound or the music of a series was built on how one particular model sounds, expect that not to be reproducible.
- For high-risk uses, ask for a human alternative. Where the music is the value in itself — a commercial, a trademark-like sound, a long campaign — the missing legal warranty is a real business risk. For background music it looks different.
More on this: AI music for commercial use, and on how much more the paid tier gives you, free and paid AI tiers — the difference here is legal, not a matter of convenience.
Common misconceptions
- "It was trained on licensed data, so I carry no risk." Whether the training data is clean is a matter between the provider and the labels. Your rights are set by the terms of service — and those exclude the legal warranty.
- "The official announcement said licensed." It did not. "Developed with our industry partners" is not the same thing, and the company is not keeping that difference by accident.
- "If I pay, the song is mine." You get what was the provider's, over what you downloaded — with no guarantee that any copyright came into being at all.
- "I can use the free version too, I just will not download it." The free tier is expressly for personal, non-commercial use.
- "The German judgment settled the question." The judgment is not final, it covers six works, and it was handed down in Germany. The merits stage of the American cases comes in 2027.
What to take away
In 2026 music AI moved into the phase where the big providers build their models together with the rightholders — while litigating with the ones they could not come to terms with. For the user that means less risk than two years ago, but not zero: the contract still pushes the responsibility onto you. The practical handhold is simple: paid tier, downloaded file, receipt kept, and the more the music itself is the value in your use, the more caution it deserves.