$1.5 billion, about $3,000 a book. A judge finalized Anthropic's piracy settlement.

A federal judge signed off on Anthropic’s $1.5 billion settlement Monday. It closes the class action that authors brought over pirated books used to train Claude, and it stands as the largest copyright recovery ever recorded.

U.S. District Judge Araceli Martínez-Olguín granted final approval in San Francisco, calling the deal “meaningful relief” for the writers and publishers whose work Anthropic pulled from pirate libraries. The number is the headline, but the shape of the deal is what other AI labs are reading closely: it puts a dollar figure on training data that was never licensed, without setting any binding rule about who wins the next case.

What the deal actually covers

The money is real, and for a copyright fund the math is unusually clean. Anthropic isn’t cutting a token check and walking; the total works out to a per-book figure that dwarfs what most infringement cases return, and nearly all of the affected titles have a named owner waiting on payment. Here’s what the deal spells out.

• $1.5 billion total, paid to authors and publishers whose books Anthropic downloaded from pirate repositories, per Reuters.
• About $3,000 per work across an estimated 500,000 books, a higher per-title payout than most copyright cases because duplicates and public-domain titles were stripped out first, NPR reported.
• 91% of the 482,000-plus covered books have already been claimed by rightsholders now due payment.
• Anthropic agreed to destroy the pirated copies it stored, so the same files can’t seed a future model.

The books came from named sources. Judge William Alsup’s earlier findings traced roughly 5 million files to Library Genesis (LibGen), 2 million to the Pirate Library Mirror (PiLiMi), and close to 200,000 to the Books3 collection, according to NPR. The suit, filed in 2024 by novelist Andrea Bartz and other writers, is where those numbers first surfaced.

Fair use survived. The piracy didn’t.

This is the distinction that keeps getting flattened, so it’s worth getting right. In a June 2025 ruling, Judge Alsup found that training an AI model on books Anthropic had lawfully bought was fair use. That part went Anthropic’s way.

What sank the company was how it got the copies. Alsup found Anthropic saved more than 7 million pirated books to a central library, a stash that existed whether or not every title fed a training run. Downloading and storing that hoard was the infringement. The settlement resolves the damages the company faced for that conduct, not the training question. Alsup gave the deal preliminary approval in September 2025; he has since retired, which is why final sign-off fell to Martínez-Olguín.

She trimmed the lawyers before approving. Plaintiffs’ counsel asked for $187.5 million in fees; the judge awarded a little over $101 million, roughly 7% of the fund, and pushed the savings back to the class. Objections that the settlement was too small, she wrote, were “not grounded in a realistic assessment of the overall risks and rewards of a trial.”

A price tag other labs are watching

No precedent came out of this. A settlement approval doesn’t rule on the merits, so it binds no other court and decides no pending case. What it does is put a number on the table.

That number matters because the piracy pattern isn’t unique to Anthropic. LibGen and Books3 turned up in training pipelines across the industry, and rightsholders in music, news, and publishing have their own suits running. A $1.5 billion recovery gives every one of those plaintiffs a reference point for what unlicensed corpora can cost. Whether any of them collects anything close is a separate fight this deal doesn’t settle.

Anthropic, for its part, wants the chapter shut. “We’re looking forward to bringing this matter to a close,” the company said. Plaintiffs’ lead counsel Justin Nelson was blunter about the milestone: “It is the largest known copyright recovery in history.”

What this means for you

If you write books, check whether you’re in the class. With 91% of 482,000 titles already claimed, the window on the rest is closing, and about $3,000 per work is real money for a mid-list author with several titles in the pool. If you build with AI, the lesson is narrower than the headline suggests: the court blessed training on legally acquired text and punished the piracy that fed it. Provenance is the exposure. Where your training data came from, and whether you can prove you had the right to use it, is now a line item with a known price attached. Don’t read this as a verdict on the bigger question of whether AI training needs a license at all. That one is still open, in courtrooms Anthropic just stepped out of.