AI Training Data Prices: What Buyers Pay per Work, Hour, Image, Token and Record (Not Compute Costs)

AI training data has no public price list: the per-unit prices on record are 21 figures stated by named parties, from USD 0.005 per street-level image to USD 5,000 per book title, and each belongs to one contract, vendor or program on one date (as of November 9, 2026). Bee Maps lists the image price; the Authors Guild described the book fee on November 19, 2024. The one per-record price in a public filing is USD 100 per delivered record, in the order form Tempus AI signed with Pathos AI on October 22, 2026.

Unit prices are one measurable part of AI training data licensing, and they come from five kinds of source: filings and signed agreements, court records, vendor list prices, buyer or program claims, and contributor pay pages. Each figure keeps its stating party and date, and nothing is averaged.

DataPacts Editorial Team checked every stated price and date on this page against its primary source on November 9, 2026; the page on how we source, date and check every figure sets out the rules.

What Counts as an AI Training Data Price, and Why Compute Costs Do Not

An AI training data price is what a buyer, vendor or program states is paid or charged for a unit of data licensed for training, not what it costs to run the computers that train a model. Searches for "AI training data cost" mostly return estimates of compute spent on training runs, a separate subject not covered here.

A figure counts as a unit price when a named party states it, the unit is countable (a work, an hour, an image, a million tokens, a record or a year) and the source is dated. Two kinds of figure fail the test:

• Contract totals without a unit. Tempus AI's 10-K for fiscal 2025 states USD 200 million in data license fees from Pathos over three years, including USD 50 million upfront.
• Per-deal payout claims. micro1 states payout tiers of USD 100K to 500K, 500K to 3M and 3M+ per company (checked 2026-11-09).

A unit price can be compared across holders; a contract total depends on volume, term and exclusivity.

Public AI Training Data Price Points, Dated

The table lists every per-unit price for AI training data that a party has stated on the record, with the date each was stated or checked. Rows are grouped by unit and ordered alphabetically by stating party within each group, as of 2026-11-09.

The word "states" marks a party's own claim about its program or price page. Contributor pay is what an individual receives for recording or submitting work, the other side of the trade from a vendor's list price. The settlement row shows "not stated" because the court record sets a fund, not a license price per work.

Price Ranges by Unit, Taken Only From the Table

Each unit's range below runs from the lowest to the highest figure in the price table, never averaged, and most units rest on one or two parties. The summary is as of 2026-11-09 and uses the 22 rows above as its only source.

Per Work and per Title: Book Licenses and a Settlement

The only per-title license price on record is USD 5,000 per book, split 50-50 between author and publisher, which the Authors Guild described on November 19, 2024 for a HarperCollins license to an AI company it did not name. The license covers opted-in nonfiction backlist titles, and the Authors Guild called the split "far too much to the publisher".

The largest book figure on record is not a license price. In Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417-AMO), the court granted final approval of a USD 1.5 billion non-reversionary settlement fund on July 20, 2026 (Dkt. 680), with first payments due on or before November 15, 2026; the fund settles claims over pirated copies. Settlements in AI copyright lawsuits over training data are tracked separately from licenses. Statutory damages under 17 U.S.C. 504(c) (2024 United States Code) run from USD 750 to USD 30,000 per work, up to USD 150,000 if willful, a court remedy and not a price.

Per Hour of Audio and Footage

Per-hour prices on record as of November 9, 2026 run from USD 3 to USD 120, and they mix two sides of the trade: what vendors charge buyers for finished datasets and what programs pay contributors for recording. SpeechData.ai, a dataset vendor, states that it sells speech datasets at USD 60 to 95 per hour of audio, with its English (US) set of 1,500 hours and 150 speakers at USD 90 per hour (checked 2026-11-09). On the contributor side, Luel's task cards pay USD 12 per hour (USD 0.20 per minute) for American English three-person conversations and USD 7.50 per hour (USD 0.13 per minute) for partner audio in Uyghur, Tibetan, Hmong, Jamaican Creole and Guarani (checked 2026-11-09).

For footage, Claru states a typical first deal of USD 10 to 120 per hour of capture, paid weekly, or USD 5,000 to 50,000 for an archive license, paid net-15 on a signed license (checked 2026-11-09). Mindrift states in a blog post of September 17, 2026 that its own egocentric video projects pay USD 6 to 15 per hour and gives a market range of USD 3 to 9+ per hour of accepted footage. A vendor's list price and a contributor's pay come from different parties, datasets and dates, so no margin can be read from them.

Per Video, per Clip and per Image

Per-item prices for video and images come from two price pages: Bee Maps lists USD 6 per AI event video and USD 0.005 per street-level image, and Luel's task cards pay USD 25 for a five-minute indoor walkthrough and USD 30.00 per CAD capture clip (all checked 2026-11-09). Bee Maps sells to buyers, so its figures are list prices; Luel's are contributor pay per task. Per-image figures for stock photo libraries appear only in press reports and are not confirmed by a party to the deals, so they stay out of the table.

Per Million Tokens, per Pull Request and per Page

Code is priced three ways on the record as of November 9, 2026: Pangea states indicative rates of USD 100 per million tokens of source code and USD 40 per million tokens of git history, Turing states USD 1,000 per pull request, and Abaka AI states a ceiling of up to USD 1,000,000 per repository. Pangea's page says the rates "can change on 30 days' notice, never retroactively" and that code owners are paid within 15 days of the buyer's payment. Turing's newsletter of March 25, 2026 gives modern codebases USD 1,000 per pull request, up to USD 10,000 per codebase, legacy codebases USD 10,000 to 100,000+, and USD 50 per organizational artifact only when bundled with a codebase. Its Project Lazarus page states a different model, revenue share with nothing upfront (checked 2026-11-09). Abaka AI's Project Reborn figure is a ceiling of up to USD 5,000,000 per company (checked 2026-11-09), not a price. Shipd, run by Datacurve, displays contributor bounties of USD 80 to 220 per coding item (checked 2026-11-09).

For documents, Specific states USD 2 per accepted page on its documents page (checked 2026-11-09). It is the only per-page price on record.

Per Delivered Record and per Year

The one per-record price in a public filing is USD 100 per delivered de-identified record, in the neurology order form Tempus AI signed with Pathos AI on October 22, 2026. Tempus AI's 10-Q for the third quarter of 2026 puts the order form at up to USD 55.0 million: USD 50.0 million for 500,000 records over five years plus USD 5.0 million for updates. A per-year price comes from a collective agreement. Kopinor and the National Library of Norway agreed NOK 45,000,000 for 2026, indexed from 2027 (s.17, signed December 19, 2025). It binds a rights collective and a national library and is not a payment by a commercial AI developer. Contract-level values sit in the tracker of AI data licensing deals with parties, dates and sources.

Who States Each Price: Contracts, List Prices, Claims and Contributor Pay

A price's reliability depends on who states it: a filed contract or court record binds a party, a list price is an offer to buyers, and a program page is the buyer's own claim about what it pays. The table sorts the 22 rows by source type, as of 2026-11-09.

Only two rows rest on a document that binds both parties to a price: the Tempus AI filing and the Kopinor agreement. Nine rows are program claims, six are contributor pay or bounties, and three are vendor list prices. Pangea places code with buyers and does not buy it, so its rates state what code owners receive. The labeling rules are set out under how we source, date and check every figure.

Units Priced in Public Without a Published Rate

Several programs price data per crawl, per use or per license but publish no rate, so the unit is known and the price is not. Examples, alphabetical by party, as of 2026-11-09:

• Cloudflare: pay per crawl, closed beta, one price per zone set by the site owner (docs 2026-07-28); fees not published.
• Newsflare: 50:50 revenue share, paid from AI clients' reports of hours used; rate not published.
• pre.dev: codebases priced per token, quoted on a call; no public figure.
• Protege: revenue share "with each use"; percentage not public.
• RSL Collective: royalties when an AI application uses the content; rates not published.
• TollBit: publishers set their own rates.
• Troveo: "There is no public price sheet" (2026-07-25).
• Warner Music Group: "all of our deals are based on variable economics" (SEC filing, 2026-03-03).
• Wirestock: 50 % commission per licensed asset in commission-based projects (2026-07-13); fixed payment amounts not stated.

Unless dated otherwise, each entry was checked on 2026-11-09.

These programs meter use instead of setting a fixed rate per unit. The crawl-based models are covered under pay-per-crawl and crawler licensing.

Per-Unit Figures That Are Not Stated Here

Several per-unit figures circulate in press and on other sites but no party to the deal or court has confirmed them, so this page names the topic and leaves the number out. Six topics, as of 2026-11-09:

• per-photo and per-image figures for stock image libraries;
• a per-minute figure for video in a stock contributor program;
• a per-page figure for individuals' work documents;
• per-work book payments, including an author's share of the HarperCollins fee as press reports it and the buyer's name;
• an hourly rate from a 2021 data labeling contract;
• per-repository baselines attributed to an unnamed platform, and code rates seen only in a search snippet.

A figure enters the price table only when a party, a filing or a court record states it.

Why a Unit Price Is Not an Offer for Your Data

A public unit price shows what one party paid or asked for one dataset on one date, and it does not transfer to another holder's data without the same rights, quality, exclusivity and buyer. The four lists below apply as of 2026-11-09.

Benefits of using public prices

• A dated anchor: the Tempus AI filing fixes a per-record price on 2026-10-22, and Pangea's code rates carry a check date of 2026-11-09.
• Units by data type: speech by the hour, code by token, pull request or codebase.

Risks

• Prices change: Pangea's rates can change on 30 days' notice.
• Sides get mixed up: Luel's contributor pay and SpeechData.ai's list price measure different things.

Downsides

• Most units rest on one or two parties.
• No party has published a per-unit premium for exclusivity.

Alternatives

• A non-exclusive license, leaving the holder free to license again.
• Per-use or revenue-share models, such as those of Protege and Wirestock.
• Collective licensing, as in the Kopinor agreement.
• Not selling or licensing at all.

A company archive mixes formats, customer contracts and personal data, so its worth depends on more than a unit price. The guide to what company data is worth to AI companies covers the methods.

Where Unit Prices Connect to Payouts, Deals, Pay and Valuation

Unit prices are one part of the market section; per-deal payouts, deal values, contributor pay and task prices each have their own page.

• The AI training data market: size, payouts and history.
• AI data partnership payouts and program terms compared: per-deal claims, dated.
• Reported payouts to data sellers: advertised versus confirmed.
• AI data licensing revenue in company filings: what listed sellers report.
• Pay for AI training work: hourly rates for expert work.
• Selling speech and audio data: the audio side of the hourly prices.
• Selling video for AI training: footage archives and capture.
• Licensing source code to AI companies: code rates in context.

Reinforcement learning environments are priced per task, not per unit of data. Those prices sit under what RL environments sell for.

How Often These Prices Are Re-checked

Vendor price pages and program terms are re-checked monthly and before any change to this table, and filings quarterly. Every dated change to a program's terms is logged in the buyer program changelog.

Frequently Asked Questions

How much does AI training data cost?

No market rate exists; public per-unit prices on record run from USD 0.005 per street-level image to USD 5,000 per book title, each stated by one party. Bee Maps lists the image price (checked 2026-11-09), and the Authors Guild described the book fee in 2024. Compute costs are a different subject.

How much do AI companies pay per hour of video?

Claru states USD 10 to 120 per hour of capture, checked November 9, 2026, and Mindrift puts accepted footage at USD 3 to 9+ per hour; no buyer has filed a per-hour license price. Mindrift's figure is contributor pay.

What is the price per image for AI training data?

One per-image price is public: Bee Maps lists USD 0.005 per street-level image; per-image figures for stock libraries appear only in press and are not confirmed. It was checked on 2026-11-09 and covers Bee Maps' own imagery.

What do AI companies pay for a book?

The Authors Guild described USD 5,000 per title, split 50-50, for a HarperCollins license in 2024; the Bartz v. Anthropic settlement is USD 1.5 billion for released claims, not a license price. The court approved the Bartz v. Anthropic settlement on 2026-07-20.

How much do AI companies pay for data?

For company data, buyers state per-deal ranges rather than unit prices, and what a dataset is worth depends on rights, exclusivity and the buyer. micro1 states tiers starting at USD 100K (checked 2026-11-09). The guide on how much company data is worth to AI companies explains the methods.