Anthropic's $1.5 billion authors settlement, about $3,000 per book, gets final approval

What was sold

Nothing was sold in the usual sense. This case is a $1.5 billion class action settlement in which Anthropic pays authors and publishers to resolve claims that it copied books from pirate sources. It is included here because it is the largest public number attached to AI training data and a reference point for what rightsholders can claim.

According to the Authors Guild, the court approved the settlement on July 20, 2026 and entered final judgment. The payment is about $3,000 per work, which the court called four times the $750 statutory minimum for ordinary infringement.¹ Notices reached rightsholders for 99.5% of the works on the Works List.¹ The page does not give a total number of works.¹

How the deal came together

Three authors filed suit on August 19, 2024. On July 17, 2025, Judge William Alsup certified the class for the piracy claims, which concern Anthropic's large-scale copying of books from illegal shadow libraries.² The parties submitted a joint statement of a potential settlement on August 25, 2025 and the settlement agreement on September 5.² Alsup granted preliminary approval on September 25, 2025, after declining to approve the deal about two weeks earlier.³

The Association of American Publishers endorsed the settlement and said it sends a message that copying books from shadow libraries or other pirate sources "has serious consequences."² The settlement requires Anthropic to destroy the files it obtained from Library Genesis and Pirate Library Mirror, and Anthropic represented that neither dataset was in the training corpus of its commercially released models.¹

Final approval came before Judge Araceli Martinez-Olguin of the Northern District of California. Class counsel first asked for 20% in fees, then reduced the request to 12.5%, or $187.5 million. The court set the award at about 6.8%, roughly $101.56 million, and withheld 10% until a post-distribution accounting is filed.¹ Service awards for the three class representatives were reduced from $50,000 to $15,000 each.¹

What went wrong or drew pushback

The court received 54 objections and comments, did not reject any on procedural grounds and ruled on the substance. One theme was works left off the Works List, and the court said claims on those works are preserved.¹ Another was that $1.5 billion is low next to the $150,000 per-work statutory maximum. The court cited trial risks, including due process concerns with very large aggregate awards, and said the settlement "provides value in prompt closure."¹

The release is narrow. It covers only claims about Anthropic's past acquisition and copying of works through August 25, 2025. Output claims, future conduct and claims involving works not on the Works List are not released.¹ The fund is non-reversionary, so unpaid fees stay with authors. Any leftover money may be redistributed to class members, and an infeasible remainder would go to charity subject to court approval.¹ No payout date has been announced.¹

Splits between authors and publishers follow default percentages, or the publishing contract, and education works are treated differently. Special Master Theodore Cheng resolves disputes among claimants.¹ The earlier proposed default for works with multiple claimants was a 50-50 split between author and publisher.³

What it means for companies with data

For a company that owns content or data, this case sets a reference point but not a price. A settlement of about $3,000 per book reflects litigation risk, statutory damages and the way the books were obtained. It is not what an AI developer would pay to license a book going forward.

The more practical lessons are about the structure. The liability was counted per work, so a large collection of identifiable items multiplies exposure. The class was defined by a list, and the court preserved claims for anything not on it. And the release was limited to past acquisition and copying, which means a settlement does not settle future use.

For a company deciding whether to license data to AI developers, the case is a reminder to be explicit about what is covered, for how long, and what happens to the data afterwards. The destruction requirement for the pirated files shows that a remedy can include deleting the data. See /resources/guide/ for contract points to check.