Generic text for AI training goes for about $0.001 per word, according to the most recent public rate, from April 2024. Licensed books work out far higher. The HarperCollins and Microsoft licence comes to about $0.067 per word over its term, if a book runs to 75,000 words. That's roughly 67 times the generic rate. No public per-token rate exists.
The rest of this page shows where each number comes from, how we converted it, and where a per-word price stops making sense. Every figure is dated. All prices are in US dollars, as published.
What is the going rate per word for AI training text?
The only public per-word rate is $0.001. Daniela Braga, CEO of the data vendor Defined.ai, gave it to Reuters in April 2024. She called it "the market rate for text".
A price per word is the amount a buyer pays for each word of licensed text. It's a simple unit. You can count it before you pay, and you can check it after delivery.
At $0.001, a 75,000-word book is worth $75. A million words costs $1,000. That's cheap, and it tells you what kind of text the quote is about. It's general text in bulk, not curated or rights-cleared work.
We looked for a newer public rate and didn't find one, as of 20 October 2026. So treat $0.001 as a dated anchor. The full set of unit prices, with sources, sits in the AI training data price table.
How do book deals translate into a price per word?
Book deals are priced per title, not per word. To compare them, you need a word count. We use 75,000 words, a common length for an adult trade book. A shorter book would mean a higher price per word.
The anchor is the HarperCollins licence. Microsoft offered $5,000 per backlist title, and authors who opted in got half, according to Bloomberg's reporting summarised by Nieman Lab. The licence was reported to run for three years.
| Source of the figure | Reported price | Per word (75,000 words) | What it buys |
|---|---|---|---|
| Defined.ai rate (Apr 2024) | $0.001 per word | $0.001 | Generic text |
| HarperCollins licence, full fee | $5,000 per title | $0.067 over the term | Three-year licence |
| Same licence, per year | $1,667 per title per year | $0.022 per year | One year of use |
| Author's share | $2,500 per title | $0.033 | Half the fee |
| Bartz v. Anthropic settlement | about $3,000 per work | $0.040 | Damages, no licence |
So a licensed book sits about 20 to 67 times above the generic rate. It depends on whether you count one year or the whole term.
The settlement row is not a price. It pays for past copying, and it gives the buyer no rights going forward.
That's the line people miss most. The $3,000 figure from the Bartz settlement gets quoted as "what AI pays per book". But nobody negotiated it as a licence. A court approved it on 20 July 2026.
Why do books cost so much more per word than generic text?
The public sources don't explain the gap in so many words. But the deal terms point to three likely reasons. Treat these as our reading of the terms, not as a buyer's stated pricing logic.
The rights are clear
A publisher can grant a licence for a whole list in one contract. The buyer gets a named rights holder and a clear term. Generic text bought in bulk often comes with weaker paperwork.
The text is edited and complete
A published book is long, edited and coherent from start to finish. That's hard to find in scraped or crowd-written text. Long, clean prose is scarce.
The volume is small
A backlist deal covers thousands of titles at most. A bulk text deal can cover billions of words. Small, specific batches usually cost more per unit than bulk.
What does news and forum text cost per word?
Nobody knows from public sources. News and forum deals are reported as yearly totals. The word counts behind them are not disclosed.
Here are the totals we can source:
- Reddit and Google: about $60 million a year, reported in 2024 (see how the Reddit and Google data deal works).
- News Corp and OpenAI: more than $250 million over five years.
- Dotdash Meredith and OpenAI: at least around $16 million a year, fixed part only.
- Amazon and The New York Times: $20 million to $25 million a year.
You can't divide those by a word count you don't have. Some deals also bundle in other things. The News Corp deal reportedly includes credits for OpenAI technology. So a per-word figure would mix up very different things. The deals database keeps each total with its source and notes.
Is anyone pricing training data per token?
Not in public. A token is the chunk of text a language model reads, often a word or part of one. API prices are quoted per token, so it's tempting to expect training data to follow.
It doesn't, at least not openly. A research paper first posted in October 2025, The Economics of AI Training Data, says tokens are traded only implicitly, through inference prices. Its table of text deals lists per-title and per-year figures, not per-token ones. The latest version, revised in August 2026, notes that billing tools have started to meter tokens directly. That covers API use, not training data.
The same paper notes that most data deals stay private. They are bilateral and negotiated case by case. That explains why so few unit prices exist at all. The glossary has short definitions of licence types and other terms used here.
Why aren't company records priced by the word?
Because the value isn't in the word count. A support ticket thread or an email chain is worth something for the decision it records, not for its length. Ten short messages that settle a dispute can matter more than a long newsletter.
So the programmes that pay for company records quote ranges per company or per dataset. They usually set a figure after looking at a sample. You can see those ranges on the published payout ranges page, sorted by headcount.
The data type pages show the same pattern by source. The email data market page lists which programmes accept email archives and what they publish. None of them quotes a price per word.
How should you read a per-word price you come across?
Check five things before you compare it with anything else. Most bad comparisons fail on one of them.
- The unit. Is it per word, per title, per year or per dataset? Convert only when you know the word count.
- The period. A three-year licence divided by three looks cheaper than the full fee. Say which one you mean.
- Whose share. The $5,000 and $2,500 book figures describe the same licence. One is the publisher's fee, the other the author's half.
- Licence or settlement. Damages for past use buy no future rights. Keep them out of price comparisons.
- The date. A 2024 quote may not hold in 2026. Look for the date the figure was checked.
That last point is how this site works. Every figure carries a source and a check date, and corrections are logged. The methodology page explains how we find, check and date each one.
What would change these numbers?
New disclosures, mostly. A filing, a court document or a published rate card would replace the per-word estimates here with real ones. Publisher annual reports are the most likely place. Some already report total AI licence revenue without a per-title split.
Until then, the honest summary is short. Generic text has one public rate, about a tenth of a cent per word. Licensed books land between 2 and 7 cents per word, depending on how you count. Everything else is priced as a total, not per word.