Anthropic's $1.5B Copyright Settlement Moves Toward Final Approval, Payment Structure Still Unconfirmed

The number is $1.5B. The approval is preliminary.

Those two facts matter in sequence. Anthropic’s copyright settlement – covering approximately 482,000 works allegedly sourced from Library Genesis and Pirate Library Mirror for Claude model training, is moving through a San Francisco federal court proceeding, but as of mid-May 2026, the court had issued preliminary approval, not final approval. The distinction isn’t procedural paperwork. Final approval is the binding event. Until it arrives, the settlement terms are subject to change, and distribution to class members hasn’t begun.

Plaintiff counsel at Susman Godfrey has estimated individual payouts at approximately $3,100 per eligible work, based on reported settlement terms. That’s a calculation, $1.5B divided by roughly 482,000 works, not a court-confirmed distribution amount. The calculation also assumes the full $1.5B represents liquid value. That assumption hasn’t been verified. The settlement’s payment structure, whether it comprises cash, compute credits, or a combination, has not been publicly confirmed. Prior AI copyright settlements have included non-cash components including compute credits; whether this settlement follows that pattern is the single most consequential unresolved question for anyone evaluating the precedent value of this case.

What the court reportedly characterized as the use of pirated training data, according to court documents, was described in terms that went beyond negligence, framing it as deliberate exploitation of creative work. That characterization matters legally, but it can’t be reproduced as a verbatim quote until the opinion text is directly accessible. The underlying allegation – that Anthropic knowingly used pirated datasets, is appropriately framed as a settlement admission rather than a court finding of fact, given that preliminary approval hasn’t yet become final judgment.

Three things remain unresolved before final approval: the payment structure (cash versus compute credits), the data destruction order referenced in concurrent proceedings, and the court’s final confirmation of the per-work distribution methodology. Each of these will shape how this settlement is cited in subsequent AI training data litigation.

The catch is that the precedent clock starts ticking before those questions are answered. Publishers and platforms negotiating AI licensing terms right now are already treating $1.5B as the settlement floor for a class of this size. Whether the actual liquid value is $1.5B or something less won’t change those negotiations, it’ll just determine how far off the floor the next plaintiff opens.

Watch for the final approval hearing date and, separately, for any SEC or court docket disclosure that clarifies the cash-versus-compute-credits composition. The full background on what final approval requires and the stakeholder positions across the broader Anthropic liability landscape provide additional context. If the settlement’s non-cash composition is confirmed via primary source, that becomes a standalone story, the headline number changes materially.

The real question is whether “landmark settlement” holds as a description if the cash component turns out to be half the figure. Compliance teams tracking training data liability exposure should hold their precedent-based risk assessments at provisional status until the payment structure is confirmed. The legal record suggests this settlement establishes that pirated training data carries nine-figure liability. It doesn’t yet confirm what nine figures means in practice.