Insights · 7 min read

How Much Do AI Companies Pay for Training Data? Reported Deals

There is no single price. There are units, and each unit has its own range.

Bearish AI · October 21, 2026

People ask how much AI companies pay for training data as if there were a rate card. There is not. Prices are set deal by deal, usually under NDA, and the units differ so much that comparing them directly is misleading. What follows is every figure we could trace to a named report or filing, grouped by what was actually being bought.

Paying per task

Epoch AI interviewed 18 people across environment startups, newer labs and frontier labs. Several agreed on a ballpark for a graded reinforcement learning task:

I’ve seen $200 to $2000 mostly. $20k per task would be rare but possible.

That is one founder’s remark, reported by Epoch. The $20,000 end applied to especially complex software engineering tasks. Compute is the reason the floor is not lower: Epoch cites an estimate of about $2,400 of compute spent per task during reinforcement learning training, which suggests cheap, low-quality tasks may waste compute.

Paying per environment or project

Epoch cites a SemiAnalysis report that replicas of websites, sometimes called UI gyms, cost around $20,000 each, and one interviewee put a high-fidelity replica of a complex product at around $300,000. Contract sizes varied with task counts: one neolab researcher had seen $300,000 to $500,000 contracts, and one founder said contracts are often seven figures per quarter or more. TechCrunch reported, citing The Information, that Anthropic’s leaders had discussed spending more than $1 billion on reinforcement learning environments over a year. That was a discussion, not a confirmed commitment.

Paying per title or per archive

• A book. HarperCollins offered authors of select nonfiction backlist titles $5,000 per book for AI training, split evenly with the author, over a three-year term, as Publishers Weekly reported.
• A shut-down company. Nearly 100 deals in a year at $10,000 to $100,000 per company were reported through Forbes and Gizmodo.
• A codebase. One founder sold theirs for about $25,000, according to MIT Technology Review.
• An airline’s internal data. Google’s $10 million winning bid for Spirit Airlines’ business data, reported by Reuters, with court approval and privacy objections still in play at the time of the reports.

Paying for a content license

The biggest numbers come from publishers and platforms licensing content on a recurring basis, not from companies selling archives:

These are scale references, not benchmarks for a typical company. A major publisher or a social platform has audiences, archives and rights that a startup’s internal records do not.

What these prices do not tell you

Three caveats keep the numbers honest. First, selection: we only see deals that someone reported or disclosed, and those skew toward large, newsworthy or distressed sellers. Second, terms: a $5,000 title license for three years, a $25,000 codebase sale and a $20 million annual license are not the same product, so a single average across them means nothing. Third, preparation: in the reports we read, suppliers scrub personal data and drop HR records, and buyers scrub again, so someone is paying for that work, whether in the price or out of it.

Where prices may be heading

Two views, both uncertain. Epoch’s reading of its interviews is that cheap, low-quality tasks risk wasting compute, which argues for paying more for quality. MIT Technology Review reports that prices for archive data will likely stabilize in the coming year, unless an AI bubble pops first. One founder told Epoch that trailing labs favor supervised fine-tuning over reinforcement learning, which, if true more widely, would change what is bought. Plan on a range of outcomes, not a trend line.

How to read all this

• Compare by unit. A task price, an annual license and a one-off archive sale answer different questions.
• Mind the source. Several figures are press estimates of deals the parties did not disclose. Epoch’s numbers come from interviews, not price lists.
• Quality sets the price within a unit. Epoch’s interviewees named robustness against reward hacking as the top quality criterion, and wanted a minimum pass rate of around 2 to 3 percent, because a task nobody can solve may be broken.
• Prices are moving. Epoch’s interviewees and compute estimates point toward paying for quality, but nobody can promise the direction of prices.

What it means for a company with an archive

The honest summary is a ladder. At the bottom are small archives sold as flat packages, in the tens of thousands of dollars at most. In the middle are checked tasks and environments, sold by the unit or by the project, in the hundreds or thousands of dollars per task and tens or hundreds of thousands per environment. At the top are content licenses held by publishers and platforms, in the millions each year. Most company archives start at the bottom, and the way up is through quality, not volume.

Most company archives will not look like a major publisher’s. Their value comes from the work being unusual and checkable: a few hundred well-built tasks at hundreds of dollars each add up faster than a pile of raw files at a flat price, and a buyer can see exactly what they are paying for. That is also why preparation matters; see /insights/raw-data-vs-prepared-data-value for the mechanics.

If you want a starting number for your own material, the free estimate at /estimate gives a conservative range. It is a starting point, not an offer. Contracts and rights should be reviewed by a lawyer.