Stop Comparing the Models. Compare Their Rights Postures.

I. Why quality is the wrong axis

Yes — the field has genuinely opened up. Suno is no longer the only serious option, and several competitors have closed the gap on the thing that used to separate them: convincing vocals.

But if you take one thing from this piece, take this: quality comparisons are the least useful way to choose. Every major platform now produces commercially usable output. What differs — enormously, and in ways that will matter to you at exactly the wrong moment — is the rights posture: what the model was trained on, who agreed to it, and what the platform actually grants you.

That’s the axis this desk cares about, and it sorts the field into three groups.

II. The three rights postures

1. Licensed training. The model is built, at least in part, on material the rights-holders agreed to license. The platform can point to named deals. ElevenLabs Music is the clearest example, having launched with Merlin and Kobalt licences covering large independent catalogs. Stable Audio has made a licensed-training claim for some time.

2. Provenance-first. The model doesn’t claim licensed training, but every output is marked and traceable. Google’s Lyria is the example: reportedly carrying an inaudible SynthID watermark on every output, with blocks on artist-voice and lyric reproduction. The bet is not “we asked” — it’s “you’ll always be able to tell.”

3. Rights granted, provenance opaque. The platform grants you commercial rights, and does not disclose what the model was trained on. Several fast-improving competitors sit here, including MiniMax and Mureka. These are the ones to read carefully, because a grant from a platform is only as strong as the platform’s own position — and a grant is not an indemnity.

The transitional state. Suno and Udio occupy a fourth position: post-settlement with some majors, still in litigation with others, and both building licensed models. They are not in group one yet. They are visibly trying to get there.

III. What each posture actually protects you from

The three categories are only useful if you know what each one buys, so here is the plain version.

Licensed training protects you from the training-data question. If a model was built on material the rights-holders agreed to license, then the most dangerous unresolved question in generative audio — what went in, and did anyone consent — has an answer for at least part of the corpus. It does not make the output yours, and it does not guarantee the licence covers your specific use. What it does is remove your record from the category of things whose foundation is disputed.

Provenance-first protects you from ambiguity, not from the question. A model that marks every output is not claiming its training was clean; it is guaranteeing that the result is identifiable as synthetic, permanently and across platforms. That is genuinely valuable and almost universally misread as a downside. Marked output is a liability only if your plan depended on nobody knowing. If your plan is the documented hybrid workflow this masthead teaches →, marking is a feature — it draws the line for you rather than requiring you to draw it.

Rights granted with opaque provenance protects you contractually and nothing else. You get a licence from the platform. That licence is a promise, and a promise is worth exactly what the promiser can back. A broad grant sitting on an undisclosed training base is a grant with an unknown floor — it may be entirely sound, and you have no way to evaluate it. For a sketch, that’s fine. For a commercial release with a supervisor’s legal team behind it, it is the weakest position on this page.

And a fourth thing none of them protect you from, which is the reason the provenance desk → exists: none of these postures say anything about your human contribution. A licensed model does not make your record human. A watermark does not document your performance. The platform’s position and your record’s position are separate questions, and only one of them is yours to answer.

IV. The comparison chart

Read this by column, not by row. The rightmost column is the one that decides whether you can safely release the result.

How to read it. Nobody occupies a perfect square. The licensed-training platforms tend to trail on raw output quality; the quality leaders are the ones with the most complicated legal histories; and the platform with the cleanest major-label position is the one most producers have never used.

And notice the pattern in the fourth column. Two of the eight mark their output. As the certification argument → holds, that is going to look less like a constraint and more like an advantage as detection infrastructure matures — because a marked synthetic file is what makes an unmarked human file legible as human.

One nuance the chart can’t hold: what “a right to use” means

Google’s training-data statement is the most carefully worded claim in this category, and it deserves reading closely rather than filing under “licensed.”

Google says Lyria was trained on materials YouTube and Google have “a right to use under our terms of service, partner agreements, and applicable law.”

That is not the same claim ElevenLabs is making, and the difference matters. ElevenLabs negotiated agreements with rights-holder organisations — Merlin and Kobalt — under which artists opt in and are compensated. Google is asserting a right derived substantially from the terms of service users already accepted when they uploaded to its platforms, plus existing partner agreements.

Both may be entirely lawful. They are not the same thing, and only one of them involved anybody being asked. A creator who uploaded to YouTube in 2015 did not opt into model training in any meaningful sense; they accepted a terms document.

This desk’s position is that “we have the right to use it” and “the people who made it agreed to this use” are different sentences, and the gap between them is precisely the territory the whole provenance argument → occupies. Google’s marking discipline is genuinely industry-leading. Its consent story is a different question, and the chart’s “provenance-first” label is meant to capture exactly that: strong on traceability, silent on permission.

The one where the terms contradict each other

This is the most important finding of the verification pass, and it is a different problem from opacity.

MiniMax publishes more than one terms document, they are scoped to different products, and they do not say the same thing. Reading its own published terms:

• The App and Web terms state that MiniMax does not claim ownership of user contributions or generated content — while granting itself a royalty-free, perpetual, irrevocable, worldwide, non-exclusive licence over that content. And the same document states that the terms permit use of the Services for personal, non-commercial use only.
• The music-creation terms address input rights extensively — requiring users to hold rights to any lyrics, audio, and particularly any voice they submit — without resolving output ownership in the same terms.
• The Open Platform (API) terms state that, as between the customer and MiniMax and to the extent permitted by law, the customer retains ownership rights in inputs and generated content — while permitting MiniMax to use both to provide, maintain, develop, improve, secure, and enforce the service.

The practical consequence for a producer is stark. A consumer-facing page may tell you that you own your output and can use it commercially. A published terms document governing the app and web product says personal, non-commercial use only. Those are different answers, and which one applies to you depends on which product you accessed and which document governs it.

This desk is not alleging bad faith. Fast-moving companies accumulate separately-drafted terms across separately-launched products, and the mismatch is more likely administrative than deliberate. But the effect on a user is identical either way, and the effect is what matters when a distributor or a supervisor asks what you were licensed to do.

The instruction, therefore, is specific: if you use this platform for anything you intend to release, identify which terms document governs your access route, read that one, and keep a copy of it as it stood on the day you generated. Question five below → exists precisely for this situation. And if you cannot establish which document governs, treat that as the answer.

The one platform offering the thing brands actually ask for

Buried in Stability AI’s own product page is the only instance of it on this list: legal indemnification, provided under its Enterprise licence.

That matters more than any feature comparison. An indemnity is a promise to defend and cover you if the rights position turns out to be wrong — which is precisely the assurance a brand’s legal team, a studio’s business affairs department, or a sync licensor is actually looking for. Every other platform on this page grants you rights. One of them offers to stand behind the grant.

Two qualifications, stated plainly. It attaches to the Enterprise tier, not the consumer plan — so a producer on a $12 subscription does not have it. And an indemnity is only worth the balance sheet behind it, which is a genuine consideration for any AI company at this stage of the market.

And one consent nuance the chart can’t carry. Stable Audio’s licensed dataset gave AudioSparx artists an opt-out. ElevenLabs’ arrangement with Merlin and Kobalt is opt-in. Both are licensed training; they are not the same standard of consent. Opt-out means your work is included unless you act. Opt-in means it isn’t included unless you choose. For a publication that treats consent as the substance of this argument rather than the paperwork around it, that distinction belongs on the record.

And one platform where the terms and the marketing diverge

Worth singling out, because it is the clearest live example of why this whole page exists.

Mureka markets its paid output as royalty-free with full commercial rights. Its own Terms of Service say something more precise: on paid tiers, ownership of the output belongs to you; on the free tier, Mureka owns all rights, title and interest, granting you a licence for lawful, internal, non-commercial use only. Those are very different products under one brand, and the difference is invisible on the marketing page.

The terms also decline to guarantee that outputs are unique or that they can be successfully registered for copyright — which is honest, and which most users will never read.

And since December 2025 the rights claim itself has been contested. A suit filed in the US District Court for the Northern District of Illinois alleges the model was trained on copyrighted recordings without authorisation, and that the “royalty-free” and “copyright-friendly” marketing is therefore false.

Two things must be said clearly. These are allegations. No court has found liability, and nobody should describe the platform as having lost anything.

But the working implication is real regardless of outcome: releasing commercially from a platform whose core rights promise is under active challenge means accepting a risk that has not been resolved. For a sketch, that is nothing. For client work, a campaign, or a release you intend to keep, it is exactly the kind of thing a supervisor’s legal team exists to find — and exactly why the eight questions below → are worth running before you commit rather than after.

V. Where the label situation stands

This section is verified against primary announcements and tier-one trade reporting. Dates and terms below are confirmed rather than inferred.

October 29, 2025 — Universal settled with Udio. The first major to settle. The agreement comprised a compensatory legal settlement plus licensing agreements covering both recorded music and publishing, and a commitment to jointly launch a new platform in 2026 powered by a model trained only on authorised and licensed music. Participation is structured as artist opt-in.

And the detail every producer should note: as part of the deal, Udio immediately disabled downloads of user-created songs — which triggered a backlash and an apparent exodus among paying users. A settlement between two companies changed what existing users could do with work they had already made, overnight. That is the single most instructive fact on this page, and Section VIII exists because of it.

November 19, 2025 — Warner settled with Udio. Settlement plus licensing across recorded music and publishing, artist opt-in, feeding the same 2026 platform. Announced roughly an hour after Warner’s separate partnership with Stability AI.

November 2025 — Warner settled with Suno. Under the announced terms, Suno will launch entirely new models in 2026 that are “more advanced and licensed,” with current models phased out.

Still live as of this writing: Sony remains in litigation with both Suno and Udio. Universal’s case against Suno continues. Claims against Suno are also being pursued in Europe, including by Denmark’s Koda and Germany’s GEMA.

Also confirmed: KLAY has been licensed by all three majors — the only AI music company to hold agreements with Universal, Sony, and Warner simultaneously. And Suno announced a partnership with BMG in August 2026, tied to its first model “developed with the music industry.”

Three pieces of analysis follow.

Licensing and litigating are parallel strategies, not opposites. Sony has licensed one platform while suing two others. Universal settled with Udio while continuing against Suno. That is not inconsistency — it is a rights-holder running two plays at once, and reading “they settled” as “it’s resolved” is reading it wrong.

The settlements are forward-looking, not retroactive. Every announced deal covers future licensed models. Universal’s and Warner’s arrangements with Udio both point at a 2026 platform trained on authorised material; Warner’s arrangement with Suno explicitly phases out current models. None of this un-trains anything that already exists.

And artist opt-in is the mechanism nobody is examining. Every one of these deals routes participation through an opt-in. Which means a signed artist’s work is licensed only if that artist opts in — and the terms on which they do, and what they receive, are between them and their label. The deal being announced is not the deal an individual artist gets.

VI. What none of the deals cover

The deal map in the previous section has a shape, and the shape is the most important thing on this page.

Every agreement flows through an aggregator of rights. A major label. Merlin, representing thousands of independent labels and distributors. Kobalt, representing a large publishing catalog. These are the entities with enough catalog to be worth negotiating with, and enough legal capacity to negotiate.

Which means the coverage map is drawn by who had a seat, not by who had music in the training data.

Three consequences worth stating plainly.

1. If you are genuinely independent, no deal on this page was made on your behalf. Not by Merlin unless you distribute through a member. Not by Kobalt unless they administer your publishing. Not by any major. The money moving through these agreements is not moving toward you, and no mechanism currently exists that would route it there.

2. Your work may still be in the models. Settlements and licences are forward-looking commercial arrangements between companies. They are not retroactive consent from individuals, and — as this desk has noted before, quoting a position stated by one of AI music’s own advocates — removing material from a catalog does not require a model to unlearn what it already took. That asymmetry is the single most under-discussed fact in this entire subject.

3. And the long-term risk is structural rather than legal. If licensed models become the industry standard — which is the direction every announcement points — then the material inside those licensed models becomes the raw material of a large share of new music. Being outside that structure means being outside the economics that follow from it, permanently, regardless of how the litigation resolves.

What to do about it, honestly: not much, individually, today. The mechanisms don’t exist yet. What you can do is make sure your own catalog is documented, registered, and legible — because when collection mechanisms for this eventually appear, they will pay the people whose ownership is provable, and they will not go looking for anyone else. That is the same argument this masthead makes about everything →, and this is one more reason it holds.

VII. How to choose, by what you’re actually doing

If you are releasing commercially, the platform’s rights posture is part of your record’s risk profile. Not a philosophical position — an operational one, sitting alongside the distribution stack → and everything else that decides where your music can live.

Read what you are actually granted, in your medium. A grant of “commercial rights” on a consumer plan frequently excludes film, TV, and game use. That exclusion is exactly where the money is, and it is exactly where nobody checks until a supervisor’s legal team asks.

Ask for the grant in writing, medium-specific, before the work — not after. The same discipline this masthead applies to every other agreement → applies here, and the platform’s marketing page is not the agreement.

Prefer platforms that mark their output — SynthID, C2PA, or equivalent. Counter-intuitive if you think of marking as a liability, but the certification argument → holds: as detection infrastructure matures, provenance is what lets a human record be told apart from an unmarked one. A platform that marks is doing part of your documentation for you.

And do not confuse a licence with ownership. A commercial licence lets you monetise the output. It is not the same as holding the copyright, and the platforms differ on this more than most producers realise.

VIII. The eight questions to ask before you release

A working checklist. Run it once per platform, not once per song — the answers change when the platform changes its terms, not when you change your track.

1. What am I actually granted — and in which media? “Commercial rights” on a consumer tier frequently excludes film, TV, and games. That exclusion is precisely where the money is.

2. Is it a licence or is it ownership? A licence lets you monetise. Ownership lets you register, sublicense, and sell. Platforms differ on this more than most producers realise, and the difference is invisible until you try to do the second thing.

3. Does the API’s terms match the consumer page’s terms? On several platforms they do not, and the API documentation is frequently the more conservative of the two. If you are building anything programmatic, read that one.

4. Is the output marked, and can I tell? Marking is a feature for a documented workflow. It is only a problem for an undocumented one.

5. What happens to material I generated under old terms? Terms change. Suno’s September 3 change → is a live example. The question is not what the terms say today — it is what they say about the files already in your sessions.

6. Does the platform’s licence survive the platform? If the company is acquired, restructured, or loses a case, what happens to the grant you relied on? Most terms address this. Almost nobody reads that clause.

7. What am I required to disclose downstream? Your distributor will ask. Your sync agreement will ask. The platform’s grant does not answer the distributor’s question, and confusing the two is a common and expensive error.

8. And the one that matters most: what is my own contribution, and can I prove it? Every question above is about the platform’s position. This one is about yours, and it is the only one that determines what tier of the Matrix → your record occupies.

IX. The honest counterpoint

1. This landscape will be wrong within months. Deals are being signed continuously, litigation is unresolved, and models leapfrog each other on a monthly cadence. Treat every specific below as a snapshot with a date on it, and check the current position before you make a decision that matters. This desk will update rather than pretend otherwise.

2. Rights posture is not the only thing that matters. If your output never leaves your own sketchpad, the licensing question is close to irrelevant, and picking a worse tool for a risk you’ll never run is its own kind of error. The posture matters in proportion to what you’ll do with the result.

3. And the uncomfortable one: none of these postures resolve the underlying question. A licence from a label covers that label’s catalog. It does not cover the millions of independent recordings that may sit inside a model trained before any deal existed — and, as this desk has noted, removing material from a catalog does not remove what a model already learned from it. The deals are real progress. They are not a full answer, and anyone selling them as one is overstating.

---

X. The last thing

Two years ago this was one question: is using this legitimate? It is now at least four — what the model was trained on, who agreed, what you are granted, and what you contributed — and they have four different answers that do not resolve into each other.

That fragmentation is not a temporary mess on the way to clarity. It is the mature state of the market. Different platforms are making genuinely different bets about how this industry settles, and several of those bets will be right simultaneously for different uses.

Which means the skill is not picking a winner. It is knowing which question you are actually facing.

A sketch that never leaves your drive faces none of them. A record going to a distributor faces the grant question. A track going into a national campaign faces all four, plus a legal team that will ask you in writing. Match the rigour to the exposure, and stop paying for either more caution or less than the work requires.

And keep the part that is yours in your own hands. Every platform on this page is negotiating its position with rights-holders who are not you. The only position in this entire landscape that you control completely is what you personally contributed and whether you wrote it down.

This comparison became the first panel of a running series. Where the models’ output can actually earn is mapped in Stop Asking If AI Music Pays. Ask Which Tier of It Pays Where. →, the coming tool-pedigree audit is argued in The Next Thing Platforms Will Audit Is Your Tools →, and the distributor gate every release passes through is surveyed in The Distributor Is the First Gate →.

Sources — Section V verified this session against primary and tier-one reporting: UMG and Udio’s joint announcement, Oct 29, 2025 · Music Business Worldwide on the UMG–Udio settlement · Music Business Worldwide on the Warner–Udio settlement, Nov 19, 2025 · Billboard’s FAQ on the UMG–Udio agreement · AP, via Barchart, on the settlement and the download removal · Copyright Alliance’s summary of the 2025 settlements and the KLAY licences · Forbes on what the settlements do and do not cover. Suno’s BMG partnership and Spotify’s identity-labelling programme were verified directly from primary announcements. ElevenLabs and Google Lyria verified this session: ElevenLabs’ launch announcement, Aug 5, 2025 · Billboard on the Merlin and Kobalt licences · TechCrunch on the opt-in structure · Music Ally on the commercial-use restrictions · Google’s Lyria 3 announcement and Lyria 3 Pro announcement, containing the training-data language quoted above · Google DeepMind’s Lyria page · Billboard’s testing of Lyria 3. Mureka verified this session from its own Terms of Service (the free-tier/paid-tier ownership split, and the absence of any uniqueness or registrability guarantee), with the December 2025 litigation reported in independent review coverage — stated as allegation throughout, with no finding of liability. Stable Audio verified this session from Stability AI’s own Stable Audio 2.0 announcement (the AudioSparx licensed dataset, its scale and composition, the artist opt-out, and the Audible Magic content-recognition partnership), Stability AI’s current Stable Audio product page (the fully-licensed-dataset claim and Enterprise indemnification), and Music Business Worldwide’s coverage. MiniMax verified this session from its own published terms: the App and Web Terms of Service (no ownership claimed over user content, alongside a perpetual licence grant and a personal, non-commercial use limitation) and the Audio Music Creation Terms of Use (input rights and voice-consent requirements). Every platform claim on this page has now been verified against a primary source. Last verified 2026-08-12.