How AI training data is priced
AI training data is priced by hour, task, dataset license, subscription, revenue share or auction. What drives each price and how to compare quotes.
AI training data is priced under six models: per hour of human labor, per task or example, per dataset license, by subscription, by revenue share and through auctions. The right model depends on whether you are buying labor, a license or software. This guide explains each model, what moves the price and how to compare quotes.
What are the six pricing models?
| Model | You pay for | Common in |
|---|---|---|
| Per hour | Expert or reviewer time | Expert demonstrations, RLHF review |
| Per task | Each example or comparison | SFT data, preference data |
| Dataset license | Rights to a fixed dataset | Text corpora, video, speech |
| Subscription | Ongoing access to updates | Data exchange listings |
| Revenue share | A cut of downstream revenue | Creator and publisher licensing |
| Auction | A competitively set price | Unique archives, procurement |
What drives the price of labor-based data?
Labor-based data follows the labor market. The profession sets the rate: lawyers, physicians and quantitative analysts cost more than general contributors. Three other factors then adjust it: the review depth (single pass or double expert review), the task length and the quality bar. A tight rubric and reviewer calibration raise cost per example and reduce rework.
What drives the price of licensed content?
License prices reflect four things: uniqueness against the public web, rights clarity, freshness and exclusivity. A clean, exclusive, regularly updated archive commands more than a static, non-exclusive one. Reported examples include about $60 million a year for the Reddit and Google license and about $250 million over five years, as estimated by The Wall Street Journal, for the News Corp and OpenAI deal. The deal tracker lists these with sources and dates.
How do you compare quotes?
Normalise every quote to the same unit and quality bar:
- Convert to cost per accepted example, not per submitted example.
- State the acceptance test and who runs it.
- Include rework, review and project-management fees.
- Compare exclusivity, term and permitted uses on licenses.
- Ask for a pilot at fixed scope before committing.
The RFP template puts these points into a document you can send to vendors.
Where do auctions fit?
Auctions help when price is unknown and several buyers or sellers compete. A reverse auction works when a buyer has a precise spec, while a sealed-bid format suits a one-off dataset. For repeatable catalogue sales, a posted price is simpler. The guide to AI data auctions covers the trade-offs.
Frequently asked questions
- How much does AI training data cost?
- There is no standard price. Human expert data is priced mainly by the expert's hourly rate, licensed content by negotiated annual or one-time fees, and catalogue datasets by subscription or license. Reported content deals range from undisclosed sums to tens of millions of dollars a year.
- Why is expert data more expensive than crowd data?
- Experts are scarce and their time has a high outside value, and expert judgments carry more signal per example.
- Are AI data prices public?
- Rarely. Most vendors quote per engagement. Reported figures come from press coverage and filings.
Editorial note. No company listed here pays for placement or has reviewed its entry. AI Data Auction is building its own auction marketplace and will label any commercial relationship with a listed company. Figures are as reported by the named source on the stated date; check the primary source before relying on them. See our methodology.