Universal and Sony Sue Suno Again Over 60,202 Recordings, Targeting v6 Distillation

TL;DR

On Friday, September 18, Universal Music Group and Sony Music filed a second copyright suit against Suno in the US District Court for the District of Massachusetts, case 1:26-cv-14275. The 45-page complaint covers 60,202 sound recordings and takes direct aim at v6, the model generation Suno launched nine days earlier with licensing deals from Warner Music Group, BMG and Believe. The legal theory is the part worth your attention: the labels argue v6 was built by distilling Suno's earlier, unlicensed models, and that training a new model on the outputs of an infringing one "does not eliminate the infringement; it launders it."

What was actually filed

The complaint landed in Boston federal court on September 18 and runs 45 pages. Warner is not a plaintiff. It left the original fight after signing its own licensing deal in November 2025, which is why this is a two-major action rather than a three-major one.

The headline numbers:

• 60,202 sound recordings alleged to have been copied without a license, which the labels describe as "only a small portion" of the works actually affected.
• Up to $150,000 per work in statutory damages for willful infringement, an arithmetic ceiling of just over $9 billion.
• Up to $2,500 per act of circumvention under the DMCA's anti-circumvention provision, roughly $150 million more on the same recording count.

This is a second suit rather than an expansion of the first for a procedural reason. Judge F. Dennis Saylor IV denied the labels' motion on August 18, 2026 to bolt 61,026 recordings onto the original June 2024 case, which covers 560 works. His reasoning was that the additions would delay a case he wants resolved on fair use, and that the labels were free to bring them separately. They did, four weeks later.

The argument that should worry anyone who distills

Strip out the music and this is a case about model lineage. The labels' framing is blunt: "V6 is not a fresh start; it is the fruit of the same poisoned tree."

The complaint alleges Suno "developed v6 using knowledge distillation and related model-transfer techniques," that the teacher models were themselves built from unlicensed recordings, and that the student therefore "also inherits the value of Plaintiffs' Copyrighted Recordings." Licensing Warner, BMG and Believe catalog for the new run does not clear the old run, because the old run is still in the room.

Think of it as a student who never opened the textbook and only ever read the teacher's lecture notes. The labels are arguing that if the teacher wrote those notes by photocopying the textbook, the student's education is still a copy. Whether US copyright law agrees is an open question, and that is exactly the point: nobody has answered it yet.

There is a second vector in the complaint that gets less attention and is arguably more novel. Suno's users spent two years picking one generated track over another. The labels say those preference signals were collected from outputs of allegedly infringing models, and therefore "embodies the expressive features" of their recordings. If that sticks, your RLHF and preference data inherit provenance problems from the model that produced the candidates, not just from the corpus.

The part that is evidence, not theory

The distillation claim is a legal argument. The DMCA count rests on something more concrete. The filing says Suno pulled audio from YouTube using yt-dlp to train v4 through v5.5, and it leans on a July 2026 breach of Suno's own code that exposed scraping from YouTube Music, Deezer and Genius.

That distinction matters for how you read the case. A court can rule against the labels on the distillation theory and still find for them on circumvention, or the reverse. They are separate counts resting on separate facts.

What Suno says

These claims remain fundamentally flawed on both the facts and the law. Suno exists so that more people can make new music. Over the past two years, we have doubled down on this goal, launching v6 in partnership with WMG, BMG, and Believe. v6 was trained on content licensed from our partners, interactions including creations and preference signals from our community, and the accumulated learnings from our team.

Read that last sentence twice. "Preference signals from our community" and "the accumulated learnings from our team" are, almost word for word, the two surfaces the complaint attacks. Legal and marketing rarely agree on what counts as a feature.

Suno's own v6 announcement is quieter than the statement. It names the three partners, says the company is retiring every previous model and moving entirely onto the v6 generation, and does not itemize the training data. No answer to this complaint has been filed yet.

Why this matters if you train anything

Suno spent two years arguing it did not need permission, then a year buying it, and got nine days of peace. The useful lesson is not about music.

Distillation is not a provenance eraser

Plenty of teams treat a clean-room retrain on synthetic outputs as a way to shed a messy corpus. This complaint is the first high-profile test of whether that works legally. If the teacher's training data is contested, a court is now being asked to say the student inherits the contest.

Preference data has provenance too

Your comparison pairs, your thumbs-up logs, your reward model inputs: if the candidates came out of a model with a disputed corpus, the labels' theory says that data is derivative. Most model cards say nothing about where preference data came from. That is about to look like an omission.

Retiring the old model does not retire the old data

Suno deprecated every previous model. The complaint is about what those models taught the one that replaced them. Deleting checkpoints does not delete lineage, and lineage is what is on trial.

Write down your teacher

If you distill, document the teacher, its training data, and its license, in the same place you document your own. Related generators like Udio are in the same legal neighborhood, and the discovery burden falls on whoever cannot show their work.

The caveats

These are allegations. Nothing here has been proven, and Suno has not yet responded in court.

The $9 billion figure is arithmetic, not a demand. Statutory damages are set by a factfinder within a range, and courts very rarely apply the willful maximum across tens of thousands of works. Treat it as the ceiling of the box, not the size of the check.

Fair use remains undecided. Judge Saylor said plainly that he wants the original case resolved on that question, and this second filing exists precisely so it does not get in the way. The distillation theory may never reach a ruling if fair use answers the question first.

Key Takeaways

• Universal Music Group and Sony Music filed a second suit against Suno on September 18, 2026 in the District of Massachusetts, case 1:26-cv-14275, covering 60,202 sound recordings.
• The core claim is that v6 was built via knowledge distillation from Suno's earlier unlicensed models, and that this launders infringement rather than removing it.
• Preference signals from users interacting with older models are named as a second infringement vector, which extends the exposure to RLHF-style data.
• Statutory ceiling is just over $9 billion at $150,000 per work, plus roughly $150 million in DMCA circumvention claims at $2,500 per act.
• This is a separate case because Judge Saylor denied adding 61,026 recordings to the 560-work June 2024 suit on August 18, 2026, and told the labels to file on their own.
• Warner Music Group, BMG and Believe have licensing deals with Suno and are not plaintiffs. Warner exited the original case after settling in November 2025.

Sources: Music Business Worldwide, Variety, Billboard, Engadget, The Next Web, Runtimewire, Suno, CourtListener