AI labs pay anywhere from a few hundred dollars for a single expert-built training task to hundreds of millions of dollars for multi-year access to a publisher's archive. For ordinary company records, the clearest public signal is Forbes' 2026 report of shut-down startups selling their Slack, Jira and email archives for $10,000 to $100,000 each.
Most data deals are private, so the public record is thin and tilted toward big media companies that disclose deals to investors. The figures below are the ones we could trace to a named source. Each comes with what it covered, because a price only means something next to what was bought.
Content licenses from large platforms and publishers
These are the deals that make headlines. They bought large archives of human writing, often with ongoing access to new content.
| Seller | Buyer | Reported figure | What it covered | Source |
|---|---|---|---|---|
| Not named | $203 million in aggregate contracts, two to three year terms | Data licensing arrangements signed in January 2024 | TechCrunch | |
| About $60 million a year, per anonymous sources | Access to Reddit content for AI | Reuters | ||
| News Corp | OpenAI | More than $250 million over five years, including cash and credits for OpenAI technology | News content from News Corp titles | Press Gazette |
Two things stand out. The biggest deals run for years, with payments spread across the term. And the sellers keep producing new content, so the buyer is paying for a flow of data as well as a back catalog. The same logic applies to an operating business, which is why we structure licenses with a monthly refresh. See recurring data revenue, explained.
Archives from closed companies
The closest public match for an ordinary business is the market for archives from startups that have shut down. Forbes reported in April 2026 that SimpleClosure, a firm that helps founders wind down companies, had handled nearly 100 sales of Slack messages, Jira tickets, emails and other records in a year, with payouts of $10,000 to $100,000 per company.
Those were one-time sales of a frozen archive with no new records to come. An operating company can sell the same kind of history and then keep licensing new records each month.
Health data
Health data has some of the largest disclosed deals, mostly from companies whose business is data. In April 2025 Tempus announced a three-year, $200 million agreement with AstraZeneca and Pathos covering data licensing and model development fees to build a cancer foundation model. Tempus reported $241.6 million of data and services revenue for 2024. Those figures bundle data with services, so they are not a per-record price.
Training tasks and environments
Labs training AI agents buy two newer products: tasks, where a model attempts a job and a grader checks the result, and environments, which are working copies of software that agents practice in. Epoch AI interviewed people who build and buy them and reported these ranges:
| Product | Reported price |
|---|---|
| One expert-built task | $200 to $2,000 (multiple interviewees) |
| A very hard software engineering task | Up to about $20,000, and rare |
| A replica of a website or app ("UI gym") | About $20,000 each |
| A high-quality replica of a complex product such as Slack | Around $300,000 (one interviewee) |
| Environment contracts with labs | Often six to seven figures per quarter |
| Exclusive versus non-exclusive | Roughly 4 to 5 times the price |
Epoch also cited a report in The Information that Anthropic discussed spending over $1 billion on RL environments over the following year.
These prices matter to companies with audit trails. A history of how people handled real situations, step by step, can be packaged as episodes that sit closer to task pricing than to bulk data pricing. Our guide on episodes, records and environments explains how one source produces all three.
Human experts paid by the hour
Labs also pay people to create data from scratch. Mercor, one of the largest suppliers, reported an average expert rate of $85 an hour as of October 2025, and its CEO told Business Insider it was paying over $1.5 million a day to contractors. Hourly work is a useful comparison: records your team already created cost a lab nothing to commission, but they would cost a lot to recreate with paid experts.
What has no public price
We found no public price lists or disclosed deal values for these kinds of company data:
- Accounting ledgers and reconciliation histories
- CRM exports
- Call-center recordings
- Insurance claims files
- Legal work product
- Semiconductor process and yield data
- Manufacturing records
Anyone quoting a precise market rate for these is estimating. Our industry price pages, such as accounting and legal, say plainly where public prices don't exist and what drives value instead.
What the figures mean for a seller
- Deals pay more when the data is hard to get elsewhere and when new records keep arriving.
- Records with outcomes and step-by-step history can be packaged into higher-priced products.
- Exclusivity multiplies price but limits the license to one buyer.
- Rights clearance is part of the price. A lab pays for data it can prove it is allowed to use.
Our guide on what makes data valuable to AI labs covers these factors in detail.
See where your data fits
Our calculator gives an estimated range for your company from your industry, systems, years of records, team size and rights. Value your data, or see how a deal works on our seller page.
Frequently asked questions
How much do AI companies pay for training data?
Public figures range from $200 to $2,000 for a single expert-built training task, per Epoch AI, to more than $250 million over five years for News Corp's license to OpenAI. Archives from shut-down startups sold for $10,000 to $100,000 each, according to Forbes.
Are these prices for raw data?
Mostly no. The large deals covered ongoing access to content, and the health deals bundled data with model development and services. Task and environment prices pay for packaged, graded work. Raw exports usually sell for less than packaged products built from them.
Why are so few deal prices public?
Data licenses are private contracts, and most sellers are private companies with no duty to disclose them. Public figures come mostly from listed companies reporting to investors or from press reports citing sources.
Can a small company get paid like Reddit?
No company should expect platform-scale figures. The closer comparison for a small or mid-size business is the startup archive market and the per-task prices for agent training, plus recurring payments for each new month of records.