The 91.3 Per Cent Signal

What the Bartz v. Anthropic Settlement Prices and What It Leaves Unpriced

The signal before the fairness hearing

The claims process in Bartz v. Anthropic PBC closed at a claim rate of 91.3 per cent. Of 482,460 eligible works in the class, 440,490 have been claimed.1 The mid-point checkpoint on 19 March 2026 sat at 54 per cent. The figure climbed thirty-seven percentage points in the final eleven days before the 30 March claim deadline.

That is not a normal claims-made settlement. The often-cited benchmark for claims-made consumer class settlements is much lower: the FTC’s 2019 study reported a median calculated claims rate of 9 per cent, excluding cases where the calculated rate exceeded 100 per cent.2 Bartz is therefore an outlier signal. It suggests that notice reached the people it needed to reach, that class members saw meaningful value in claiming, and that the practical force of adequacy objections is reduced.

The fairness hearing is no formality. Judge Martínez-Olguín has moved the hearing from 23 April to 14 May 2026 and has permitted timely objectors to be heard.3 The arithmetic, though, is already doing heavy work in the plaintiffs’ favour.

The distinction that forced the deal

The piracy theory, not the training theory, is what Anthropic is paying $1.5 billion to settle. The distinction matters.

On 23 June 2025, Judge Alsup granted partial summary judgment holding that the copies used to train Claude and its precursors were fair use, and that digitising purchased print books was also fair use. He refused, however, to treat Anthropic’s pirated central-library copies as protected by the same analysis.4 The reasoning was analytical. Training on copyrighted texts to synthesise a generative model, on the court’s view, is transformative in the manner a human reader is transformative when they read widely to learn to write.

Library Genesis and Pirate Library Mirror were not merely acquisition channels for training. On Judge Alsup’s analysis, they were the source of a permanent, general-purpose corpus of infringing files retained on the company’s servers. The court treated that corpus as a separate use requiring its own justification. Statutory damages for wilful infringement run up to $150,000 per work.5 For a class of roughly half a million registered works, the theoretical exposure ran into the tens of billions. The $1.5 billion settlement is priced against that retention exposure, not against the training use itself.

Anthropic settled because the piracy and central-library issue created live, class-wide statutory damages exposure, not because Judge Alsup had rejected fair use for training as such. The implication for the broader AI copyright docket is structural.

The settlement in numbers

The gross fund is $1.5 billion. Requested deductions total roughly $208.6 million, producing a net distribution fund of approximately $1.29 billion.6 The base payout per claimed work, before interest accrual, comes to $2,931.62. Claims will be adjudicated and distributions calculated by approximately 11 June 2026, with payment commencement expected no earlier than late autumn 2026.7

Anthropic has placed $300 million into escrow with further tranches due after final approval, before 25 September 2026 and before 25 September 2027.8 The staged funding is standard in settlements of this magnitude and is one reason final approval is not a formality. The court needs to be satisfied that the funding schedule is secured against Anthropic’s balance sheet and against foreseeable corporate change. It is unusual for a copyright defendant to pay on this scale over this period, and the court will want the tranche mechanics scrutinised in open court.

The 12.5 per cent fee request

Class counsel have moved the attorney fee request from 25 per cent of the fund to 12.5 per cent, equal to $187.5 million.9 Expenses add $2.78 million, a cost reserve of $18.22 million sits alongside and service awards of $150,000 are distributed across the three named plaintiffs at $50,000 each.10

The fee arithmetic is the part of the settlement most likely to draw favourable attention at the fairness hearing. Class action fee percentages often sit materially higher than 12.5 per cent: the Ninth Circuit’s common-fund benchmark is commonly stated as 25 per cent, and Authors Guild describes class action fees as ranging around 30 per cent. A 12.5 per cent request is uncommonly low in absolute terms and markedly low in relative terms against a fund of this size. The step-down reads as a strategic move to pre-empt a common adequacy objection and to narrow the surface area available to objectors who want to attack the economics.

The unsealed objections

The court ordered a substantial block of objections unsealed in early April 2026, including dockets 425, 539-546, 549-552, 564-569, 584-585, 588-589 and 593-612.11 Authors Alliance later noted that, as of 20 April 2026, several of the listed objections still appeared to remain sealed, including dockets 544, 596, 598, 600, 601 and 602. Named objectors include Professor Lea Victoria Bishop and an objector identified as Esquivel.

The grounds fall into five families. First, scope. The class as certified excludes foreign and non-US-registered works. Objectors put the excluded population at more than two million books. Second, quantum. Statutory damages run up to $150,000 per wilfully infringed work. A settlement offer averaging roughly $3,000 per work represents a reduction of some 98 per cent against the statutory ceiling. Third, registration. Group registrations are being treated as single compensable works even where they collapse forty or more titles. Fourth, split. The allocation mechanism permits publishers to claim approximately half of the per-title payment before the author sees any of it. Fifth, precedent. Some objectors argue the settlement sets a cheap liability template that the next defendant will try to replicate.

None of those objections is obviously frivolous. The first is a structural line-drawing choice that reflects the plaintiffs’ registration-based theory of infringement rather than a settlement compromise. The second and third go to adequacy and are the familiar territory of fairness review. The fourth reflects longstanding author-publisher tension over book-revenue allocation and the fifth is policy rather than law. Judge Martínez-Olguín will weigh each on the Rule 23(e) adequacy standard, but the 91.3 per cent claim rate cuts against all of them as a matter of record. A class that overwhelmingly elects to claim has, on a practical reading, told the court that the deal was worth taking.

Kadrey, Meta and the limits of the parallel

Bartz leaves Judge Alsup’s fair use holding on training untouched. Kadrey v. Meta Platforms, Inc. reached a similar bottom-line result on training, but on a materially different and heavily caveated record.12 Judge Chhabria granted Meta summary judgment on fair use, while making clear that the ruling did not establish a general rule that Meta’s use of copyrighted works to train language models was lawful. It reflected the plaintiffs’ failure to develop the evidential record needed to defeat fair use.

The safer proposition is therefore narrower. On the current Northern District of California authorities, training remains a fair use defence available to model developers. It is not a settled rule. The next result will depend on the record: provenance, market harm, output evidence and whether plaintiffs can move beyond the evidential gaps that weakened Kadrey.

The route that remains most clearly open is the one Anthropic lost on. Show that the model was trained on, or built on top of, data the defendant did not lawfully acquire, and pursue retention and reproduction theories rather than the training theory in isolation. That is the litigation template visible in the dockets now moving through 2026 and 2027. It runs through discovery into acquisition provenance rather than through argument about the fair use factors alone.

OpenAI, Midjourney and the live dockets

The next wave of cases does not mirror Bartz on the facts. The consolidated OpenAI copyright litigation in the Southern District of New York, before Judge Sidney Stein, now includes a 5 January 2026 order affirming Magistrate Judge Ona Wang’s direction that OpenAI produce the full 20 million-log ChatGPT sample.13 The Disney, NBCUniversal and Warner Bros. Discovery actions against Midjourney and MiniMax, commenced between June and September 2025, raise output infringement on a visual-corpus theory that did not feature in Bartz at all.14 Those cases go to whether the model’s outputs reproduce copyrighted expression in a way the training cases did not need to decide.

The Bartz settlement does not dispose of any of those theories. It establishes two things only. It confirms that where a defendant lifted pirated corpora from LibGen or PiLiMi, liability is real and pricing can reach roughly $2,931 per registered work. It confirms, by not disturbing Judge Alsup’s holding, that on the current Northern District of California authorities training on legally acquired material remains a fair use defence available to the model developer. Everything beyond those two points is open.

The fork

The analytical question for 2026 and 2027 is whether follow-on plaintiffs can demonstrate that legal acquisition was a cover for scraped or pirated data. If they can, $2,931.62 per registered work becomes a reference point for damages discussions, not the ceiling. If they cannot, the Bartz and Kadrey defendants’ record on training holds on its narrow facts, and the plaintiffs’ bar will need to pivot to output infringement theories of the kind now pending in the Midjourney and MiniMax dockets.

That is the fork the AI copyright bar now faces. The 91.3 per cent claim rate tells the court that eligible rightsholders overwhelmingly elected to participate in the settlement economics. It does not tell future courts what to do with the next model. Future courts will ask where the training data came from, whether any of it was pirated and whether any of the outputs reproduce protectable expression. None of those questions has a settled answer. All of them are the subject of active litigation.

The settlement, if approved after the 14 May hearing, will draw a line under Anthropic’s class-wide piracy exposure for the covered works. It will not draw a line under the AI copyright question. The industry will read $1.5 billion as the price of piracy. The plaintiffs’ bar will read $2,931.62 as a reference point for registered works. Both readings are right and both are partial. The rest of the question is in discovery.

Footnotes

1. Authors Guild, Anthropic Settlement Update: 91.3 Percent of Books Claimed (17 April 2026) https://authorsguild.org/news/anthropic-settlement-update-91-percent-of-books-claimed/

2. Federal Trade Commission, Consumers and Class Actions: A Retrospective and Analysis of Settlement Campaigns (2019) (median calculated claims rate of 9 per cent across claims-made consumer class settlements, excluding cases where the calculated rate exceeded 100 per cent).

3. Authors Alliance, Bartz v. Anthropic Settlement Update: New Date and Time for the Fairness Hearing and Unsealed Objections (14 April 2026, updated 20 April 2026) https://www.authorsalliance.org/2026/04/14/bartz-v-anthropic-settlement-update-new-date-and-time-for-the-fairness-hearing-you-can-join-online-and-unsealed-objections/

4. Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.), Order on Partial Summary Judgment (23 June 2025) (Alsup J).

5. 17 U.S.C. § 504(c)(2).

6. Authors Guild (n 1).

7. ibid.

8. ibid.

9. PYMNTS, Anthropic Copyright Settlement Lawyers Cut Fee Request to $187.5 Million (20 March 2026) https://www.pymnts.com/cpi-posts/anthropic-copyright-settlement-lawyers-cut-fee-request-to-187-5-million/

10. Authors Guild (n 1).

11. Authors Alliance (n 3).

12. Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal.), Order on Cross-Motions for Partial Summary Judgment (2025) (Chhabria J). The court granted Meta summary judgment on fair use while making clear that the ruling did not establish a general rule that Meta’s use of copyrighted works to train language models was lawful and reflected the plaintiffs’ failure to develop the necessary evidential record.

13. Reuters, OpenAI loses fight to keep ChatGPT logs secret in copyright case (3 December 2025) (reporting Magistrate Judge Wang’s December 2025 production order); Jones Walker, OpenAI Loses Privacy Gambit: 20 Million ChatGPT Logs Likely Headed to Copyright Plaintiffs (January 2026) (reporting Judge Stein’s affirmance of 5 January 2026); In re: OpenAI, Inc. Copyright Infringement Litigation, S.D.N.Y.

14. MBHB, AI News Roundup: Anthropic settles copyright infringement lawsuit for $1.5 billion, Warner Bros. Discovery sues Midjourney for copyright infringement (2025); Disney and Universal v. Midjourney, Inc. (commenced June 2025); Warner Bros. Discovery v. Midjourney (commenced 4 September 2025); Disney, Universal and Warner Bros. Discovery v. MiniMax (commenced 16 September 2025).