This post is going out early as a thank-you to paid subscribers. The full piece opens up to free subscribers next Tuesday. I’m grateful you’re here, and grateful for your commitment to rethinking what education can be. Your support is what keeps this work going.
Somewhere in the last couple of weeks, I lost track of which AI model launched when. There was a new video generator, then another one, a frontier text model with a decimal-point version number, and image models that leapfrogged each other before anyone had finished writing the comparison post. I can no longer tell you, without checking, which one came out first. I know that whatever came last is supposed to be better than what came before, because the benchmark line always goes up. But that does not seem to matter as much anymore.
Call it model fatigue. The release cadence has outrun our ability to follow it.
One launch out of that blur, however, deserves more attention than the rest. On September 9, 2026, Suno shipped its v6 family of music models. The audio it generates is impressive, but what interests me more is where the model came from and what it represents. As far as I can tell, Suno v6 is the first major creative model released under a public licensing framework negotiated with companies that control rights in the music. That does not settle every argument about generative music. It does, however, remove the most immediate provenance objection and shift the others into new territory.
And for those of us who teach, there is another, maybe even harder, problem behind the legal one. What justification remains for making something yourself once the provenance objection and the obvious AI tells both recede into the background?
In today’s post, I want to trace how a copyright lawsuit turned into a training license, separate what is known about the data behind v6 from what is still opaque, describe what the model sounds like now, and then ask what we should tell students once provenance and technical flaws are no longer enough to dismiss the output.
How a lawsuit turned into a license
The story starts on June 24, 2024, when Sony, Universal, and Warner sued Suno in federal court in Massachusetts. The Recording Industry Association of America coordinated the case. The complaint alleged that Suno had copied copyrighted recordings at scale, without permission, to train its models.
The evidence was unusually concrete for a case of this kind. In tests of Suno’s output, the labels found producer tags—the spoken audio watermarks commercial producers stamp on their beats—embedded in generated tracks. A tag like “CashMoneyAP” would be difficult to explain unless unedited commercial recordings had entered the training data. The labels also documented outputs that closely tracked specific songs, from Chuck Berry’s “Johnny B. Goode” to Mariah Carey’s “All I Want for Christmas Is You,” along with synthetic vocals that sounded a great deal like Bruce Springsteen and ABBA.
Suno’s defense was the one pretty much every generative AI company has used at this point: fair use under 17 U.S.C. § 107. The argument is that a model does not copy protected expression in order to redistribute it. Instead, it extracts statistical relationships between chord progressions, rhythms, and timbres, and then generates something new. And this is roughly what a human musician also does by listening to a lot of records.
I have some sympathy for the fair-use argument as a matter of principle, and I have written before about how the “it’s just like a human learning” analogy fits what these systems actually do. However, given what was at stake, it always seemed unlikely that the fair-use question would reach a definitive judgment. Statutory damages for willful infringement run up to $150,000 per registered work. The labels that are still suing have since identified more than 60,000 recordings, creating potential legal exposure in the billions.
The way Suno got its training audio made the legal risk worse. In May 2025, Suno told the labels in discovery that it had downloaded audio from YouTube with open-source tools, including yt-dlp. That September, the labels asked the court for permission to add a separate claim under the DMCA’s anti-circumvention provision, 17 U.S.C. § 1201, arguing that those tools circumvented YouTube’s technical protections. A fair-use argument would not by itself resolve that claim.
It is therefore not surprising that on November 25, 2025, Warner Music Group settled with Suno and announced what both sides called a partnership. The settlement agreement itself has never been made public, so what follows comes from the two companies’ announcements and from press reporting.
Warner licensed Suno to build new models with recordings and compositions it controls. While the exact financial terms were not disclosed, Suno says a share of its revenue now goes to its licensing partners. Artists and songwriters signed to Warner must opt in before Suno can use their names, images, likenesses, or voices in its features. And a recognizable vocal clone of a Warner artist may appear only with that artist’s approval.
The partnership also came with product restrictions. Suno agreed to retire its v3, v4, and v5 models, the ones trained on scraped data, once a licensed replacement was ready. Free-tier users lost the ability to download audio at all, and paid users got monthly export caps. Both measures were intended to curb the flood of AI tracks onto streaming services and the resulting dilution of the royalty pool. And in a piece of the deal that reads more like an acquisition than a settlement, Suno took over Songkick, Warner’s concert-discovery platform, along with behavioral data on its millions of registered users.
Those terms materialized in the v6 launch. The download restrictions are now live, and the old models are being retired as v6 rolls out. The model itself appears to be the core deliverable of the partnership.
What is actually in the box
Suno describes v6 as a family. The flagship v6 is aimed at paying subscribers and tuned for predictable, polished results. A sibling called v6-wild loosens the constraints for producers who want stranger textures, and a distilled v6-mini powers the free tier, where you can generate and stream but not download. All of them accept more than text. You can hand v6 a video clip, a still image, or a voice memo, and it will generate a score for it, matching cuts, mood, and pacing. It also supports localized edits, such as rewriting a chorus while leaving everything else unchanged.
