Start with the multiplication, because that is the first thing a finance partner will ask for. Take the plaintiffs' own phrase, tens of thousands of compositions [2], at its floor of 20,000, and multiply by the $150,000 per work they are asking for [8]. That is $3bn [1], twice the $1.5bn Anthropic agreed to pay authors and publishers to settle a separate class action over pirated books, reported as the largest US copyright settlement at the time [10]. Run the division the other way and that settlement is worth exactly 10,000 works at the same per-work rate [2]. BMG's separate case, at 493 compositions [11], tops out around $74m, roughly 2.5 percent of the floor Sony and Warner are working from [3].
Statutory maxima are ceilings. Cases settle underneath them. None of those figures is a forecast. What they size is the reserve your model vendor may spend the next two years arguing over, while your roadmap quietly assumes stable pricing and stable terms.
The reproduction allegation is the part that reaches into your product rather than Anthropic's data pipeline [4]. A case built on output means the exhibit is a transcript: a prompt someone typed, and what came back. You do not have to have trained anything to generate one of those.
Teams tell themselves users treat a text feature narrowly: summarizing their own documents, rewriting their own drafts. Users treat anything that produces prose on request differently. They ask for the thing they cannot write themselves, which in the lyrics case means the next verse, in full, of a song somebody else owns.
Two axes sort this, and neither of them is how much the feature gets used. First, whether the feature asks a model to produce third-party creative text at length, or only to transform text the user supplied. Second, whether you retain and inspect outputs, so you would know which one happened.
Long-form production with no output inspection is the expensive corner, because you can neither size the exposure nor show anyone you tried. Long-form production with inspection at least lets you cap length or route the request to a licensed source. Transformation with inspection is cheap. Transformation without inspection holds up until a user finds the prompt that turns your summarizer into a lyric printer.
Switching vendors does not move the question either, since OpenAI has faced similar allegations from The New York Times and Encyclopedia Britannica [12]. Anthropic had not commented publicly when Digital Trends published its account of the filing [9], and the naming of Amodei and Mann as individual defendants [5] is worth a customer's second read, because it aims the claim at who approved what rather than only at a balance sheet.
The narrow version of the work: for each feature that puts model output in front of a user, decide whether a transcript of it would read badly as an exhibit, then find out what your vendor contract says in writing about that specific case. Provenance, once confined to somebody else's slide about responsible data sourcing, now carries a per-unit price attached to it.
