Suno's v6 rollout pulls every older model off the platform

Memorization lives in a model's weights, a fact about specific parameter files rather than about the company that trained them. In late July the Munich Regional Court I found Suno liable for copyright infringement, ruling that six musical works are reproducibly embedded in the version 3.5 and 4 models ^[10]. That is evidence about two specific parameter files. It says nothing about v6, and the reverse holds too: retiring a probed model removes the artifact while leaving the underlying question open.

Codeveloper is the load-bearing word here, left without a definition by Suno. Suno credits Warner Music Group, BMG and Believe as codevelopers of v6, according to the-decoder's account of CEO Mikey Shulman's blog post ^[1], while giving no detail on which catalogs were used or how much material went in ^[5]. That description fits a fully licensed training corpus. It also fits a fine-tune on label-supplied stems sitting on an unlicensed base, or a review role with no training contribution at all. The safeguards cited at launch were already published in August: drop artist names from training metadata, check uploaded audio and lyrics for unauthorized use, and hold a download policy against mass distribution of songs on streaming platforms ^[7]. Only one of those three addresses what the model saw during training; the other two govern what users put in and take out.

Shulman's engineering claim is that the new models are faster, more expressive and higher quality than their predecessors ^[15]. No measurement is offered, and the baseline needed to test it is the thing the rollout withdrew ^[3]. For a relative quality claim to transfer to your material you need the old model in the same account, taking the same prompts, with the same seed discipline. The last major generation was v5 in September 2025, with the interim v5.5 adding voice cloning in March ^[4]. Anyone who spent those months building a prompt library against v5 output is now calibrating against something they cannot diff.

The money is worth doing by hand. Suno raised $400 million in June at a $5.4 billion valuation ^[11], so the round sold roughly 7.4 percent of the company ^[17]. Set that valuation against the $300 million of annual recurring revenue reported in February ^[12] and it is about 18 times revenue ^[18]. The same revenue floor across more than two million paying subscribers works out near $150 per subscriber per year ^[19]. Both inputs are stated as minimums, so treat that per-seat figure as a floor rather than an estimate.

What the licensing arrangement delivers to someone shipping tracks is a route to market: songs made with the partner model are to be distributable through Believe and TuneCore ^[13], with per-artist offerings on an opt-in, paid basis promised next ^[14]. The scope of the underlying license is legible to the three labels, and to the lawyers who asked the US District Court in Massachusetts to seal the exact size of the training data on the grounds that competitors could infer the development approach from it ^[6]. Multimodal input and part-level text editing are real capability gains, and the editing model is well chosen: you can replace a chorus or a single lyric line without regenerating the track ^[16]. That leaves the person clearing that track for release without a document to check the license against.