On 9 September, Suno launched a family of three models: v6 as the main version, v6-wild for exploration, and v6-mini available for free. It is the first time the company has shipped several models simultaneously. They are presented as developed with the music industry — Warner Music, BMG and Believe. The previous models were withdrawn immediately. Universal and Sony have not signed and are continuing their legal action.
We have covered the question of how rights holders are paid and what artists are going through. Here is the outcome of a story we had been following for months, and it is more ambiguous than it appears.
What really changes
The novelty is not technical, it is legal.
Until now, these tools were trained on corpora whose origin was not documented, which produced the lawsuits we have covered. This launch marks a shift towards licensed inputs, negotiated with catalogue holders.
The chronology is instructive. Warner settled its dispute in November 2025 by committing to licensed models for 2026. BMG signed on 12 August 2026, with voluntary opt-in for its artists and a settlement covering prior use of its catalogue. Believe and TuneCore signed on 8 September — the day before the launch, four months after blocking their tracks on the platform.
In other words: these agreements are not the fruit of a conversion, they are exits from litigation. The lawsuit produced the licence.
What the company does not say
One point deserves to be highlighted because it is systematically absent from the coverage: the company does not detail what was actually used for training.
Announcing models developed with industry partners does not mean the corpus is entirely licensed. It may mean that some catalogues were added, that some were removed, or that an agreement retroactively covers prior use.
The distinction is not rhetorical. A model trained exclusively on authorised content and an earlier model regularised by a settlement are not the same object, and only the first formulation would really justify the label.
Despite these agreements, the legal situation has worsened. Universal Music and Sony Music have not signed and are pursuing their claims. Canadian collecting society SOCAN filed an action on 2 September. And a Munich court ruled against Suno in the GEMA case in July — a case we had been following for months and whose European outcome was widely awaited. The company is therefore launching its licensed models in a litigation environment heavier than the one it faced a year ago. Signing with three partners does not protect you from those who have not signed.
The immediate withdrawal of the old models
One detail has concrete consequences for users: the previous models disappeared from the selector on launch day.
It had been announced — the agreement with Warner explicitly provided that existing models would be withdrawn when the new ones arrived, and an email had given notice on 5 September with compensation in credits.
The consequences are no less real. Tracks already produced remain in libraries, but the vast majority of guides and prompting advice published online now describe models that can no longer be selected.
This is exactly what we described in our article on depreciation: a tool accessed as a service exists only as long as someone else decides to maintain it. Here, the decision came from a legal agreement.
What changes for those who create
The announced features go in a direction consistent with what we observe everywhere: editing rather than generation.
Modifying a section of an existing track in plain language, while preserving the rest. Combining elements from several sources in one request. Isolating an element and building on it.
It is the same shift we noted for ChatGPT Images 2.5. Generating impresses; editing precisely is what you need when you are working.
Trials carried out on launch day report an observation that should temper a great deal of advice: the gap between two generations from the same request can exceed the effect of a deliberate change of wording. In other words, a good share of the prompt recommendations circulating measures noise rather than a real effect. It is a methodological reminder that holds well beyond music: to know whether a formulation is better, you have to compare it across several trials, not one.
What to take away
This launch illustrates the direction we anticipated: the lawsuits were not aimed at preventing training, they were aimed at setting its price. We wrote as much about Sony and Warner's complaint against an AI lab: every settlement establishes a price, and that price serves as the reference for the next one.
What remains open is the question that interests artists most, and which none of these agreements answers publicly: how much flows back down to creators. An agreement between an AI company and a catalogue holder settles a dispute between two companies. What the writer of a track used in training receives is a different negotiation, less visible, and whose terms nobody publishes.
>>