Anthropic's $1.5B author settlement gets the judge's stamp - only 350 opt-outs

In plain terms - A US federal judge has given final approval to Anthropic's $1.5 billion settlement with the authors who sued over Claude being trained on pirated books. Only 350 authors opted out of the class - a tiny fraction - which means the deal essentially binds the whole plaintiff class.

Contexte

The case (Bartz v. Anthropic) targeted Anthropic's use of a large corpus of pirated books during Claude's training, sourced from shadow libraries. In June 2025 the same court had already ruled that fair use covered training on lawfully acquired books but not on pirated copies - a split ruling that made the piracy claim the operative liability. The $1.5 B figure emerged from settlement talks that followed.

Les faits

• The presiding federal judge (Northern District of California) approved the settlement on 21 July 2026 (AP wire, confirmed by Ars Technica and The Verge).
• The class covers authors whose works appear in the identified pirated corpora. Anthropic will pay approximately $1.5 billion, with per-book allocations to be administered by a claims process.
• Approximately 350 authors filed formal opt-outs - a fraction of the class of tens of thousands.
• The settlement resolves the piracy claim; it does not adjudicate the broader question of whether training on lawfully purchased copyrighted books is fair use - that split ruling from 2025 stands.

Analyse

The low opt-out count is the story. For most rights-holders, the calculus was: individual litigation would cost more than the settlement share pays out, and the fair-use half of the ruling means they'd probably lose on the biggest legal question anyway. The class action mechanic worked as designed - it aggregated small stakes into a coherent lever.

For Anthropic, this is expensive but bounded. $1.5 B is roughly a quarter of what Meta and OpenAI have committed to spending on training runs in a single year at 2026 scale. Bounded liability with a clean legal exit is what a growth-stage frontier lab actually wants.

The precedent matters more than the money. It's now settled - literally - that using shadow libraries as training data is a payable event. Every frontier lab now has a number to negotiate against.

Scénarios

• Base case (65%) : other labs settle similar class actions in the $500 M - $2 B range within 18 months.
• Structured (25%) : industry converges on a compulsory-license style clearing scheme for training data, brokered by publishers.
• Tail (10%) : Congress or the EU codifies a training-data licensing regime - the harder path.

Implications

For a founder deciding what to train on: the piracy shortcut has a price sheet now. For rights-holders: the class action mechanism is validated. For policy: the framework will emerge from private settlements, not statutes.