AI Content Licensing Deals Tallied: Who Paid, and What the Terms Show

Only one AI content licensing deal has its price printed in a securities filing. Reddit told the SEC in February 2024 that it had signed data licensing arrangements worth $203.0 million in aggregate. We tallied 14 AI licensing deals with any public number attached, and only 5 rest on a figure the seller or buyer stated itself. The rest of the market price for training and retrieval rights is built from leaks.

The short answer: AI companies have paid individual publishers and platforms roughly $10 million to $66 million a year for content licences, based on the few terms that are public. Reddit's filing is the hardest number. News Corp, the New York Times and Dotdash Meredith fall inside that band. Per book, licences run near $5,000 per title.

This post is written for founders and directors who own a content or data asset and are deciding whether to license it, and at what price. That reader needs comparables, not commentary, so the ledger comes first.

What counts as a disclosed AI licensing deal

A licensing deal here means a contract where an AI developer pays for access to text, images or video. It may pay for training, for retrieval into live answers, or both. Training rights cover building a model. Retrieval rights cover showing or summarising current content inside a product like ChatGPT, Alexa or Meta AI.

Most of these contracts are confidential. A number becomes public in one of three ways, and they are not equally reliable. We graded every deal on which route its number took.

Grade A: the company, a filing or a court said it. The number appears in a securities filing, an earnings call, a results release or a court order. Reddit's S-1, IAC's disclosures on Dotdash Meredith, and Wiley's earnings calls sit here. These figures are still incomplete, because companies disclose aggregates and minimums, but someone is accountable for them.

Grade B: press reports citing people familiar. The number was reported by an established outlet, usually The Wall Street Journal, Bloomberg, Reuters or the Financial Times, citing unnamed sources. The companies declined to confirm. Grade B is citable. It is also how most of the market's price signal was formed, which is the uncomfortable part.

Grade C: announced with no number. The parties confirmed a deal exists and said nothing about money. That is the majority of AI publisher deals by count. We included one Grade C entry, the Meta news bundle, because it shows how the newest deals are structured. We excluded dozens of others that carry no number at all.

The ledger: 14 AI licensing deals and their terms

Where a figure is a minimum, a range or a contested reading, the row says so. Annualised values divide a reported total by the reported term, and that is our arithmetic, not the company's.

Grade A means the figure came from the company, a filing or a court. Grade B means press reporting from unnamed sources. The Reddit and Google row and the Reddit aggregate row likely overlap, because the filing does not name its buyers. Two News Corp rows involve a WSJ report about its own parent company.

Read the Grade column before the money column. The two largest press figures, News Corp with OpenAI and News Corp with Meta, were both reported by The Wall Street Journal, which News Corp owns. That does not make them wrong. It does mean the seller's own newsroom set the public anchor price for news licensing, and you should weigh that.

AI publisher deals on an annual basis

Licensing terms are quoted in different shapes: totals over five years, annual minimums, per-title fees and revenue pools. To compare them, convert each to a single year where the reported terms allow it. Where they do not, leave the deal off the chart rather than guess.

The ceiling is set by platforms, not newspapers

The two highest bars belong to Reddit. A platform with a live, high-volume archive of human conversation commands at least as much as the largest English-language newspaper group.

News Corp's two deals sit just below. If both WSJ reports are accurate, News Corp alone draws up to roughly $100 million a year from two AI buyers.

The floor is about 1% of a publisher's revenue

Two deals at the bottom of the band have been sized against the seller. The Amazon payment to the New York Times is nearly 1% of the Times's 2024 revenue, per the WSJ. An analyst put the OpenAI payment to Dotdash Meredith at about 1% of that publisher's revenue, according to Engadget's report on the IAC disclosure.

My read: 1% of revenue is the number that matters most for a publisher deciding whether to sign. It is meaningful money and nowhere near a replacement for the search and social traffic AI answers may displace. Anyone selling a licence as a new business line should be shown that ratio first.

Training data licensing priced per book

Book deals are the only part of the market quoted per unit, which makes them the cleanest comparables for anyone pricing a catalogue.

The HarperCollins agreement with Microsoft pays $5,000 per title, split evenly with the author, according to Sherwood's report of the Bloomberg story. It covers selected nonfiction backlist titles, and authors who decline are left out. The Authors Guild objected that a 50/50 split gives the publisher too much for a training licence.

Why a licence costs more than a settlement

A licence buys things a settlement does not. It buys a clean provenance record, curated titles the buyer selected, and agreed output limits. Reporting on the HarperCollins deal said outputs are capped at 200 consecutive words or 5% of a book. A settlement buys only release from past claims.

The Anthropic settlement releases claims about past acquisition and copying through 25 August 2025 only. It does not cover output claims or future conduct. So the $3,000 figure is the price of a past liability, not a forward right to use anything.

Academic publishers price by catalogue, not by title

Academic publishers have not disclosed per-title rates. They report totals. Wiley reported $40 million of AI licensing revenue in fiscal 2025, up from $23 million, including a single $18 million agreement with a multinational tech customer. Publishers Lunch reported $49 million for fiscal 2026, with recurring AI revenue of $8 million inside that.

That recurring line is the number I would watch. In fiscal 2026 most of Wiley's AI revenue still came from one-off dataset sales. Bloomberg Law similarly described Informa's 2024 AI income as $75 million of non-recurring data access. A one-off sale is a windfall. It is not a business.

Reddit's licensing numbers, the only ones in a filing

Reddit is the reference point because it is the only seller that put contract value into a securities filing. Its S-1 said that in January 2024 it entered data licensing arrangements with an aggregate contract value of $203.0 million and terms of 2 to 3 years. It expected a minimum of $66.4 million of that to be recognised in 2024, as TechCrunch reported from the filing.

The filing did not name the buyers. Reuters reported, citing one source, that the Google contract was worth about $60 million a year. The S-1 risk factors added that "substantially all of the contract value associated with our licensing revenue is derived from one of our partners".

Reddit's 2025 Form 10-K reports total revenue of $2,202.5 million, up 69%. It says other revenues increased "as a result of content licensing agreements executed in 2024 and 2025". The same risk factor now reads "two of our partners" instead of one, consistent with the second reported buyer, OpenAI.

The $66.4 million 2024 minimum was about 5.1% of Reddit's $1,300.2 million in 2024 revenue. Advertising was $2.1 billion of 2025 revenue, per the auditor's report. Licensing is material to Reddit's story and small in its income statement.

The most useful sentence in the filing is not the $203 million. It is the admission that almost all contract value comes from one or two buyers. There are perhaps five or six buyers with the budget to pay eight figures, and they all know what the others paid.

How licensing prices compare with litigation outcomes

Licensing prices are negotiated in the shadow of the courts. If a buyer believes training is fair use, the rational licence price falls toward zero. If courts disagree, it rises toward statutory damages. Three outcomes now bound that range. Our breakdown of the copyright cases that could reprice training data covers the legal reasoning.

Divide $1.5 billion by the roughly 482,000 eligible works and you get about $3,100 per work. Then compare it with the HarperCollins licence at $5,000. Settling after the fact cost Anthropic less per title than licensing up front costs Microsoft. The $1.5 billion figure frightens people because of its total, not its unit price.

I disagree with the common read that this settlement proves licensing is the cheap option. Per unit, it suggests the opposite. What it proves is that pirated acquisition is the expensive mistake, which is a narrower lesson.

The Third Circuit's Ross decision is the first federal appellate ruling on fair use in AI training, per IPWatchdog's report. Commentators note it distinguished generative systems from Ross, which returned existing passages. That narrowing matters, but it hands specialist publishers whose content feeds a competing tool a stronger hand than they had a month ago.

Where this tally is weakest

A ledger that claims to set a market price should state where it cannot.

Most rows are not verifiable from primary documents

8 of 14 rows are Grade B. We opened the reporting for each, but the underlying contracts are confidential. If the WSJ's sources overstated the News Corp figures, the top of the band moves down by half.

Survivorship shapes the sample

Deals leak when someone benefits from the number being public. A seller wants a high figure out. A buyer rarely wants any figure out. So the reported deals may skew high, and the dozens of undisclosed deals may sit well below the band. The opposite is possible too: OpenAI equity and compute credits, which News Corp's deal reportedly includes, may make headline cash look smaller than the real consideration.

The units do not match

Annualising a five-year total assumes even payments. Reddit's $66.4 million is a minimum across buyers. Wiley's figures are fiscal-year revenue across several buyers. Disney received a licence fee nobody disclosed alongside a $1 billion equity investment going the other way. Treat the annualised chart as a rough band, not a price list.

Pricing your own content licence from these comps

If you own a content or data asset, the ledger gives you three anchors and one warning. Our view on whether to let crawlers in at all sits in the robots.txt decision for AI crawlers, and that decision comes before pricing.

Sell retrieval, not just training

Training rights are a one-time event. Once a model has learned from your archive, the buyer has little reason to renew. Retrieval rights to current content renew because the content goes stale. The Amazon, Meta and Perplexity structures all pay for display and summaries, and that is the side of the contract I would negotiate hardest on.

Perplexity's model is the clearest example of recurring pricing. It committed a $42.5 million initial pool and 80% of Comet Plus subscription revenue to publishers, paid by traffic, citations and agent use. The pool is small. The per-use structure is the right shape.

Citation and traffic are a separate question from licence fees. How AI answers choose which sources to cite is covered in the playbook on Reddit and Perplexity citations, and a licence does not guarantee either.

The News Corp deal with OpenAI reportedly included credits for News Corp's use of OpenAI technology. Disney took the opposite position, investing $1 billion in OpenAI alongside its licence. When money flows both ways, the headline fee is not the price. The pattern is familiar from the circular deals across the AI supply chain, and it deserves the same scepticism.

Frequently asked questions

Where to start if you own the content

Before any buyer calls, write down which grade of comparable you are anchoring on. If your target price comes from a Grade B leak, say so in your own model, and test what happens to your case if that number is half as large.

Then split your asset in two on paper: the archive you would sell once, and the current feed you would license by use. Price them separately. The archive sets your one-time windfall. The feed is the only part that can still be paying you in 2029.

Related reading