A federal judge approved Anthropic's $1.5 billion settlement with authors over pirated training books on July 21, 2026, with only 350 writers opting out of a class covering roughly 500,000 works. The deal, first announced in 2025, pays statutory damages of about $3,000 per book for titles Anthropic downloaded from pirate libraries to train early Claude models, before it built the licensing pipeline it now runs for training data acquisition.

The number that matters for every other lab is the opt-out count, not the settlement figure. 350 authors out of a class this size means the statutory-damages math Anthropic accepted, roughly $3,000 per work regardless of that book's actual commercial harm, was cheap enough and fast enough that 99.9% of eligible plaintiffs took the check instead of suing separately. OpenAI is still fighting a parallel copyright claim from the New York Times in a trial where discovery is ongoing, and Wired reported this week that OpenAI's pre-release models were behind a security breach of Hugging Face's evaluation infrastructure, the model-hosting platform labs use to benchmark unreleased systems. Every lab still training on scraped or pirated text now has a priced precedent to weigh against continued litigation risk: Anthropic's per-book rate is the number opposing counsel will cite first.