The Anthropic settlement is final: 3,000 dollars a book and the opt-outs

Where the case ended

This month Judge Araceli Martinez-Olguin of the Northern District of California granted final approval to Anthropic's 1.5 billion dollar settlement with a class of book authors. Objections from affected authors were overruled. Some authors and publishers had already opted out to pursue their own suits, and those will continue separately. The class action itself, filed in August 2024 by Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, is over.

The arithmetic that everyone will carry forward is about 3,000 dollars per book plus interest, spread across a class of roughly half a million authors. The claims process had its own calendar: opt outs and objections were due February 9, 2026, reinclusion of previously excluded works by March 9, and claims by March 30. Those dates are past, so the pool is now fixed.

How a fair use win became a 1.5 billion dollar loss

It is easy to forget that Anthropic mostly won this case on the law. In June 2025 Judge William Alsup granted summary judgment that training on legally purchased books was fair use. What he refused to excuse was the sourcing. The court found Anthropic had used over seven million pirated library copies, and it sent that question to trial. The settlement is the price of not having that trial.

The sequence matters for anyone reading the settlement as a verdict on training. It is a verdict on acquisition. A lab that bought and scanned the same books would, on Alsup's reasoning, have paid nothing to this class. The 3,000 dollar figure is what it cost to have downloaded them instead.

The judge who said no first

The settlement was announced in August 2025 and rejected in September by Judge Alsup, who objected to shortcomings in the settlement details that would be forced down the throat of authors. His stated concern was that works from the Pile, the dataset containing the unlicensed material, remained unresolved. The deal that Martinez-Olguin has now approved is the reworked version. She is also the judge who handled Tremblay v. OpenAI, where in February 2024 she dismissed most claims and kept only the direct infringement count over training data, so she came to this with a narrow view of what these cases are about.

We do not know from the public record what changed between the rejected and the approved versions beyond the process around the Pile works. That is the piece of this story we would most like to read a proper account of, because it decides whether authors whose books were never individually identified got anything at all.

What 3,000 dollars establishes

A settlement is not a ruling, and 3,000 dollars is not statutory damages. It is the number two parties agreed to under the threat of a trial where the statutory range ran far higher per work. Still, it is now the only large, public, per-work price for pirated training text, and prices like that get cited. Expect it to appear in every negotiation between a publisher and a lab from here on as the floor for what unlicensed use costs after the fact.

The more interesting effect is on the licensing market that did not exist when the books were downloaded. If a book is worth 3,000 dollars in damages for one company's training run, a publisher will ask for something in that region up front for a license, and a lab will point out that licensed use was never the problem. The cases that opted out will test whether a jury sets a different number, and until they conclude, 3,000 dollars is the anchor.

What we would watch

Three things. First, the opt out suits, because a single jury verdict on the same facts could reset the anchor in either direction. Second, whether other pending author cases against OpenAI and Meta, including the Authors Guild suit and the Silverman and Kadrey claims, settle on a similar per-work basis or fight the fair use question that Alsup answered in Anthropic's favour. Third, whether any lab publishes a provenance audit of its pretraining corpus voluntarily, because after this settlement the cost of not knowing what you trained on has a number attached.