Sony and Warner attach a $150,000 per-work price to Anthropic's training corpus
The per-work figure in this week's complaint is small enough to look modest and large enough that, multiplied across a corpus, the provenance of training data stops being a legal footnote and becomes a budget line.
The Product Desk

What happened
- Sony Music Publishing and Warner Chappell Music sued Anthropic this week in federal court in Northern California, in a filing first spotted by Music Business Worldwide.
- The publishers ask $150,000 for each work infringed across tens of thousands of works, which they say could push damages well past a billion dollars if the case succeeds.
- The complaint alleges Anthropic used BitTorrent to pull lyrics and sheet music for hundreds or more compositions, naming Livin' On a Prayer, September, Hallelujah and others.
- It argues that Anthropic treats last September's $1.5 billion authors settlement as the cost of doing business, and that the sum was too small to deter further infringement.
- An earlier suit brought by Universal Music Group, Concord and ABKCO in October 2023 was seeking up to $3 billion as of January 2026, on comparable allegations about song catalogs.
Why it matters
- cost A published per-work rate lets finance model something that used to sit in the risk register as a paragraph, and the multiplier is the number of works nobody in the building can actually list.
- constraint A corpus assembled by torrent is hard to enumerate after the fact, so any team that wants a defensible number is pushed toward licensed or documented sources even where scraped data trains better.
- decision Buyers of vendor models now have to decide whether their indemnity covers how the training corpus was acquired, not only what the model emits at inference time.
- precedent With a rate in public, the next rights holder to come forward can size a demand before discovery, so future claims arrive priced rather than exploratory.
Most vendor security questionnaires have a field for training data sources, and most answers to it contain no nouns. This week's complaint gives that field a unit price: Sony Music Publishing and Warner Chappell Music want $150,000 for each work they say was infringed [3]. Once a rate like that is in public, the only variable left is how many works someone can show you hold.
So do the division. Anthropic's authors settlement, the largest copyright settlement in American legal history [6], works out to 10,000 units at the publishers' rate [10]. Sony and Warner put their own count in the tens of thousands [3]; read that as 20,000 to 99,000 works and the same rate yields $3 billion to nearly $15 billion [12]. The floor of that range is what Universal Music Group, Concord and ABKCO are already seeking as of January 2026, in a case they filed in October 2023 [7][11].
Teams shipping a generative feature tell themselves provenance is the model vendor's problem and the contract handles it. What these filings describe happens earlier than any contract: acquisition of the corpus, allegedly over BitTorrent, down to individual lyric and sheet music files for songs including "Hallelujah" and "Livin' On a Prayer" [4]. Sony and Warner call it "one of the largest and most blatant ongoing thefts of intellectual property in history" [2]. The reason a rate can be quoted at all is that copies are countable, while output similarity is an argument.
The test that travels to your own stack has two axes. First, whether you can list the works in each corpus your feature depends on, the works and not the source. Second, whether a named party pays when that list turns out longer than you thought. Listed works with a named payer is a licence you can put in a budget. Listed works with nobody named is a negotiation you still owe. Unlistable works with a named payer means your exposure sits on someone else's balance sheet, and their filings are worth an afternoon. Unlistable works and nobody named is the box these complaints are aimed at.
PCMag, which reported the filing, notes that its own parent Ziff Davis sued OpenAI in April 2025 over the training and operation of its systems [9], so nobody narrating this is a neutral party. That is a reason to price the risk rather than argue it out in a standup. If the answer to the first axis is a vendor's name instead of a list, the box has already been picked for you.
What to watch
- Whether the Sony and Warner complaint is amended to state an actual work count, which converts "tens of thousands" into a ceiling.
- Whether any Anthropic settlement with music publishers is struck on per-work terms rather than a lump sum, which would leave a quotable rate behind.
- Whether model vendors begin offering enumerable corpus documentation, or indemnity that names acquisition as well as output, in enterprise contracts.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Insufficient
- Hype gap+25
- Incentives80
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Sony Music Publishing and Warner Chappell Music filed a copyright lawsuit against Anthropic in federal court in Northern California earlier this week; the filing was first spotted by Music Business Worldwide.
- [2]
The suit accuses Anthropic of "one of the largest and most blatant ongoing thefts of intellectual property in history" and of a "brazen campaign of illegally torrenting, scraping, and downloading copyrighted works on a massive scale" to build its AI models.
- [3]
The publishers are claiming $150,000 per work infringed, across tens of thousands of works, meaning potentially well over a billion dollars in damages if the case succeeds.
- [4]
The filing alleges Anthropic used BitTorrent to download the lyrics and sheet music to hundreds or more copyrighted musical compositions, including Livin' On a Prayer, Great Balls of Fire, Ramblin' Man, September and Hallelujah.
- [5]
In September 2025 Anthropic agreed to a $1.5 billion settlement of a class action brought by a group of authors and publishers whose books it allegedly used to train Claude.
- [6]
That $1.5 billion settlement is the largest copyright settlement in American legal history.
- [7]
An earlier suit filed in October 2023 by Universal Music Group, Concord and ABKCO alleged comparable misuse of their song catalogs to train Claude, and as of January 2026 those publishers are seeking up to $3 billion in damages.
- [8]
The new filing says Anthropic "clearly considers that to be just the cost of doing business given that its entire business model continues to be built on copyright theft," and that $1.5 billion is "not a large enough settlement to deter infringing conduct."
- [9]
PCMag discloses that its parent company Ziff Davis filed a lawsuit against OpenAI in April 2025, alleging infringement of Ziff Davis copyrights in training and operating its AI systems.
- [10]
At the publishers' asking rate of $150,000 per work, Anthropic's $1.5 billion authors settlement is equivalent to 10,000 works.
- [11]
At $150,000 per work, 20,000 works comes to $3 billion, the figure the Universal Music Group, Concord and ABKCO case is now seeking.
- [12]
If "tens of thousands of works" means 20,000 to 99,000 works, the $150,000 rate implies $3 billion to about $14.85 billion.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- pcmag.comAnthropic Hit With New Copyright Lawsuit From Sony, Warner | PCMag
1 article · August 30, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Training Data ProvenanceFollow
- Music Publishing RightsFollow
- AI Legal Liability CostsFollow
- AI Copyright LitigationFollow