Anthropic $1.5B Copyright Deal Nears Payouts as AI Lawsuits Mount

The settlement is final and the money is starting to move

The largest copyright recovery in U.S. history has reached the payout stage. On July 20, 2026, U.S. District Judge Araceli Martínez-Olguín granted final approval to the $1.5 billion class settlement in Bartz v. Anthropic and entered final judgment. The case covers 482,460 books.5 The settlement administrator expects the first payments to go out on or before November 15, 2026, for works where the claimants agree on how to split the money.6 Rights holders are now hearing about it locally as well. A Tufts anthropology professor recently learned that his first book was among the works Anthropic downloaded from pirate sites.1

The headline numbers hold up, but they are not everything. The deal does settle a costly mistake. It does not settle whether AI companies may train on copyrighted work, and Anthropic has described the case in a way that supports its own position.

How a three-author suit became a $1.5 billion bill

Novelist Andrea Bartz and authors Charles Graeber and Kirk Wallace Johnson filed suit in August 2024. They alleged that Anthropic had used pirated books to build its Claude models.20 The turning point came in June 2025, when Judge William Alsup split his fair-use ruling. Training the models on the books counted as fair use. So did digitizing print books that Anthropic had bought. Downloading and keeping pirated copies in a permanent central library did not.5 Alsup wrote that Anthropic had no right to use pirated copies for that library.4

Class certification changed the math. A complaint from three writers became a case covering close to half a million works, and statutory damages can reach $150,000 per work. Commentators put Anthropic's theoretical exposure above $70 billion.2 Other estimates ran into the hundreds of billions if a jury had found willful infringement at a December trial.20 Anthropic settled instead, and Alsup granted preliminary approval in September 2025.6

The settlement is narrow. It covers only Anthropic's past acquisition and use of the listed pirated works before late August 2025.2 It does not give Anthropic a license to use the works in the future. Claims about infringing outputs from Anthropic's models are also excluded.3 Anthropic must destroy the files it downloaded from Library Genesis and Pirate Library Mirror.3

What the court approved, and where accounts differ

Martínez-Olguín took over the case after Alsup retired. She wrote that the deal gave the class substantial benefits given how novel the claims were, and that a loss at trial would have left authors with nothing.11 She overruled all 53 objections.23 She pointed out that the payout of roughly $3,000 per work is four times the statutory minimum.19 Only 350 class members opted out. She refused most late opt-out requests, including one from entertainment lawyer and author Donald Passman, whom she found inexcusably late.19

The biggest part of the ruling concerned legal fees. Class counsel first asked for $300 million. They later lowered the request to $187.5 million, after Alsup questioned the original figure.6 The court awarded about $101.56 million, roughly 6.8% of the fund, and cut each lead plaintiff's service award from $50,000 to $15,000.6 Coverage generally treats the fee cut as a check on class-action lawyering. One tracker reports that the fee award has been appealed.6

The reported figures vary in small ways. Per-work estimates include "about $3,000"7, about $3,10011, and the $3,351.39 cited in Passman's filing.19 The claims rate is reported as 91.3%18 and as 92.77% after an update at the hearing.6 The gap reflects which point in the claims process each outlet measured. It also matters that the per-work amount is a gross figure, to be divided among authors, publishers and other rights holders.14 Claim notices sent in early September showed many authors that a publisher or literary agency was claiming part of a book they had expected to be paid for in full.14 Larger publishers stand to collect tens of millions of dollars.12 Disputes over how to split each payment are now the most likely source of delay.

Anthropic is paying in installments. It paid $300 million after preliminary approval. A second $300 million was due within five days of final approval. Two more payments of $450 million each are scheduled for September 2026 and September 2027.1811

Some summaries get the basic facts wrong. One tracker says the case was an Authors Guild class action that settled in "early 2026."26 The court record shows a deal announced in August 2025 and approved in July 2026.6 One major newspaper also referred to the judge as "he."13 The court-record timeline should be preferred over secondary summaries.

Both sides claim the precedent

The sharpest disagreement in the coverage is over what the case means. Anthropic's deputy general counsel, Aparna Sridhar, called the case a landmark showing that training AI on books is fair use.7 The Association of American Publishers took the opposite view. It said the court refused to accept downloading from pirate sites as an efficiency for the infringer.12 Bartz called the settlement a first step toward accountability for what she described as Big AI's theft.13

Neither framing is fully accurate. Because Anthropic settled, the fair-use ruling will never be reviewed on appeal and is not binding precedent.16 Its practical significance is economic. The settlement puts a price on past training-data liability, roughly $3,000 per book, and gives corporate risk managers a figure to work with.9 Lawyers advising companies that use AI make the same point. The deal priced inputs alone and left output claims open.25 The case did not decide whether AI companies may train on books. It made the source of the books expensive to get wrong.

The litigation is spreading

The settlement also created new lawsuits. Authors who opted out filed a complaint in May 2026 seeking individual statutory damages rather than class relief.21 Opt-outs led by Pulitzer winner John Carreyrou are suing Anthropic, OpenAI, Google, Meta, xAI and Perplexity.22 Anthropic filed a consolidated motion to dismiss on September 24, with a hearing set for December 17. It is also seeking dismissal of a separate suit on statute-of-limitations grounds.24 These cases will show whether the opt-outs bet correctly that they can win more than about $3,000 per work.

Other lawsuits against Anthropic go beyond books. Music publishers have cited Anthropic's admission that at least one Claude model was trained on a dataset containing lyrics to at least 100 of their works.22 Motions to dismiss a second lyrics suit, which also names Dario Amodei, are fully briefed.24 Courts have treated similar facts differently. In Kadrey v. Meta, training claims were dismissed on fair-use grounds, while claims about Meta allegedly redistributing pirated books through torrents went forward.22 That pattern matches Bartz. In both cases, how the books were acquired and passed along created more legal risk than the training itself.

The legal theories are also changing. After the Supreme Court's decision in Cox Communications v. Sony, OpenAI argued that contributory infringement claims brought by news publishers can no longer stand.23 The judge in the OpenAI litigation then refused to let The New York Times and Daily News add new contributory-liability theories against Microsoft.21 On authorship, the Supreme Court's refusal in March to hear Thaler v. Perlmutter leaves purely AI-generated work ineligible for U.S. copyright.27

A legal-tech case sets the first appellate rule

The most important appellate ruling so far comes from legal research, not books. On September 29, the Third Circuit affirmed the lower court in Thomson Reuters v. Ross Intelligence. It held that training a competing legal research tool on 2,243 Westlaw headnotes was not fair use.23 The court said ease of copying does not justify it the way necessity might, and found that Ross did not need the headnotes to reach the unprotected legal information beneath them.23 Ross involved non-generative AI, but observers had expected the ruling to influence every pending generative AI case.29

For legal-tech vendors and law firms adopting AI tools, the result is significant. Building a product on a competitor's curated editorial content is now high-risk at the appellate level, even as a single district court has called training on lawfully acquired books transformative. Analysts expect cleaner data sources, licensing, and narrower domain-specific models to follow.9 Advisers also warn that the risk extends to companies that only use AI in products, content and code, not just the developers who build models.25

What to watch

The next major test is New York Times v. OpenAI, where cross-motions for summary judgment ask the court to apply the full four-factor fair-use test to a record that includes evidence of models reproducing articles verbatim.28 A win for the Times would sharply change the landscape. A win for OpenAI would strengthen the current trend toward settling and licensing.26 In the Meta book cases, the court refused an early appeal and said a final judgment would soon allow the issues to reach the Ninth Circuit together.21

For now, the Anthropic settlement mainly sets a price. Using pirated material costs about $3,000 per work. Whether training itself requires payment remains undecided. With most of the money paid in during 2026 and 2027, the next stage depends on appeals courts, not class-action settlements.