Rung 05 DataSwitch rungClose
The Price of Training Data
Early StageEvery other rung of this KB can point at a price: a GPU-hour, a watt, a million tokens. The data that a model learns from has mostly been free to take, so it has had no price to point at. The Bartz v. Anthropic settlement gives it one — not a market price, but a penalty price set in court. Anthropic agreed to a $1.5 billion fund covering roughly 500,000 qualifying books, which works out to about $3,000 per title, over 7 million book copies it had downloaded from the pirate libraries LibGen (June 2021) and PiLiMi (July 2022) (Authors Guild summary).
Read carefully, that number is a ceiling, not a going rate. It is what one company paid to settle for books it took without paying. A licence negotiated in advance would sit under it — which is exactly why a court-set penalty is useful: it bounds the negotiation. The $3,000 is also a minimum before fees and costs, and what each author actually receives depends on how many valid claims come in against the ~500,000 estimate.
The unit matters as much as the total. Divide the same $1.5bn by the 7 million copies downloaded instead of the ~500,000 titles covered and you get about $214 per copy. The settlement uses titles. The $214 is arithmetic done in the close-read only to show how much "the price of a book" depends on the denominator; it is not a figure anyone agreed to and should not be cited.
The negotiated side: what a licence cost
Until 2026-09-24 this page had only the penalty. Reddit's IPO filing gives the other side: data bought, in advance, through a deal. In January 2024 Reddit signed data licensing arrangements with an aggregate contract value of $203.0 million over two to three years, and expected to recognise a minimum of $66.4 million of it in 2024 (Reddit S-1). Spread over the stated terms, that is roughly $68–102 million a year — this KB's arithmetic on the filing's two figures, not a number the filing states. Substantially all of the contract value came from one partner, whom the filing does not name.
The two prices are not the same unit, and that is the lesson:
| Bartz settlement | Reddit licence | |
|---|---|---|
| What was bought | Peace over books already taken | Continuing access: an API plus quarterly bulk transfers |
| Unit | Per registered book (~$3,000) | Per year of access (~$68–102M, all contracts) |
| Set by | A court, after the fact | A negotiation, in advance |
| Buyer count | One defendant | Substantially one partner |
| As-of | 2025-09-25 | Contracts January 2024; filed 2024-02-22 |
Neither source says how many tokens its data amounts to, so the two cannot be put on one scale. What they do show together is that both prices of training data this KB holds were set by essentially one buyer each — which is why neither is a market rate yet.
How it reaches the token price — and what this KB cannot yet say
On the path from a token price to a task price, data is the input at the very start: a one-off cost paid before a single token is served, then spread across every token the model ever produces. The settlement prices that input for one kind of text — registered, commercially published books — got one way: unlicensed. The Reddit filing prices another kind — a living forum corpus — got the other way: licensed, per year. Neither says how much of any model's training set its data is, or how many tokens the cost is spread over, so neither can be turned into a cost per token. That conversion is the open question this rung exists to answer. The supply side — how long free public text lasts before buyers have to pay — is on the data stock.
Key Claims
- $1.5bn fund, ~500,000 qualifying titles, ~$3,000 per title minimum. Preliminary approval 25 September 2025. Evidence: moderate (the terms are court record, but this KB holds them via the Authors Guild's summary, not the settlement agreement itself) (Authors Guild)
- The eligibility gate narrows the unit sharply. A work qualifies only with an ISBN or ASIN and a US Copyright Office registration made within five years of publication, and registered either within three months of publication or before Anthropic's download (10 August 2022). The price is for registered, commercially published books, not "text" in general. Evidence: moderate (Authors Guild)
- This is the price of unlicensed acquisition, not of training. See acquisition versus training. Evidence: moderate for the ruling's structure; weak for how far it generalises
- A negotiated licence: $203.0M aggregate contract value, two-to-three-year terms, at least $66.4M recognised in 2024. Signed January 2024, from substantially one partner (unnamed in the filing). The partner being Google, and "~$60M a year", come from press reporting and are [unverified] here. Evidence: strong for the figures (sworn SEC disclosure); weak for any reading of them as a market rate (one buyer) (Reddit S-1)
- The licence created a revenue line from almost nothing. The 2024 minimum ($66.4M) is more than four times all of Reddit's "other revenue" in 2023 ($15.2M), and that line also includes subscriptions. Evidence: strong (filing figures; the comparison is arithmetic) (Reddit S-1)
- It does not generalise to web-scale pretraining data. One jurisdiction, one defendant, two shadow libraries, and a gate that excludes much published work. It is a datapoint, not a market rate. Evidence: the close-read's own limitation (Authors Guild)
Benchmarks & Data
| Term | Value | As-of | Source |
|---|---|---|---|
| Settlement fund | $1.5 billion | 2025-09-25 | Authors Guild |
| Qualifying titles (approx.) | ~500,000 | 2025-09-25 | Authors Guild |
| Payment per title | ~$3,000 minimum, before fees and costs | 2025-09-25 | Authors Guild |
| Copies at issue | 7 million (LibGen, PiLiMi) | 2025-09-25 | Authors Guild |
| Claims deadline | 30 March 2026 | 2025-09-25 | Authors Guild |
| Final approval hearing | 14 May 2026 — outcome not in this source | 2025-09-25 | Authors Guild |
| Reddit licensing, aggregate contract value | $203.0 million, 2–3 year terms | 2024-01 (filed 2024-02-22) | Reddit S-1 |
| Reddit licensing revenue expected in 2024 | ≥ $66.4 million | 2024-02-22 | Reddit S-1 |
| Reddit "other revenue", full year 2023 (licensing + subscriptions + user economy) | $15.2 million | FY2023 | Reddit S-1 |
Open Questions
- Was final approval granted, and when? This source gives only the hearing date. Secondary reporting has put approval in July 2026; that is unverified and deliberately not cited. Resolve from the docket before any published claim turns on it.
- What does licensed data cost now? The Reddit price is from January 2024, and its two-to-three year terms run out between January 2026 and January 2027 — the renewal price is the live number. Reddit's quarterly filings no longer break licensing out (see frontier).
- How many tokens is a book worth, and over how many served tokens is its cost spread? Without both, the $3,000 cannot become a cost per token.
Related Concepts
- Acquisition versus training — which act the $1.5bn actually priced
- The data stock and when it runs out — why free text stops being enough
Backlinks
Pages that reference this concept:
Related concepts
Referenced by (3)
Other pages in the base that lean on this one.