What training data costs, from filings and studies

How much training data costs depends on how it is obtained. Public filings and published studies give dated figures for licensed content, human labels, model-made labels and the preparation of free data. None of the filings reviewed here discloses a price per record, token, image or hour of licensed content [1, 2, 4, 5, 6].

The figures here are about data. One published estimate of what frontier models cost to train, revised in February 2025, counts hardware, energy, cloud rental and staff, and lists "the acquisition of data for model training" among the costs its methods leave out [13].

How much does training data cost?

Each figure carries its date and its evidence class.

A contract value is what buyers committed to one seller over a term, a revenue line is what one seller recognized in a year, and a cost per label is what one team paid, or estimated, for one task. None of them is a price list.

Licensed content: contract values and revenue lines

Reddit's Form S-1 of 22 February 2024 reported data licensing arrangements entered in January 2024 "with an aggregate contract value of $203.0 million and terms ranging from two to three years", and an expected minimum of $66.4 million of revenue in 2024 [1]. Spread evenly, that is about $68 million to $102 million a year (derived) [1]. The filing names no buyer and does not split the value by use.

Shutterstock's 10-K for 2025 reports $203.3 million of Data, Distribution, and Services revenue, up from $137.3 million in 2023 [4]. The line also covers Giphy and Shutterstock Studios, so it is wider than training data [4]. Wiley's 10-K for the year to 30 April 2026 reports "Research AI licensing revenue" of $33.1 million, against about $11 million a year earlier, and "Learning AI licensing revenue" of $16.0 million, against about $29 million [5]. CuriosityStream's 10-K for 2025 reports delivering "hundreds of thousands of hours of video and audio" to technology companies, with no price per hour [6].

The Copyright Office's report on generative training, listed on 8 November 2026 only in its pre-publication version of 9 May 2025, records one commenter's suggestion that rightsholders "can negotiate licenses that forgo up-front payments or traditional royalties in exchange for later shares in revenues", and the Office itself writes that compensation based on a percentage of revenue or profits "may be an attractive alternative for smaller developers" [7, 36].

One court record prices acquisition without a license. A class settlement between a model developer and authors and publishers, given final approval on 20 July 2026, provides approximately $3,000 per work before costs and fees from a $1.5 billion fund [8, 9]. Its release states that it is not "a license to torrent, scan, or train AI models on any copyrighted works" [9]. The figure is a payment for past conduct. This article is research and draws no legal conclusion.

Data annotation cost for human labels

Three studies report what human labels cost or would cost: one paid its annotators and two estimated the price [18, 24, 25]. Wang et al. priced human labels at $0.11 each in 2021, estimated from Google Cloud Platform's crowdsourced labeling rates [24]. Ding et al., from interviews with professional annotators, put 3,000 sentiment labels at $221 to $300 in 2023, about $0.07 to $0.10 each (derived) [25]. The AlpacaFarm team (2023) paid annotators a median $21 an hour, for a one-time cost of $3,000 for preference labels on 10,000 instructions, about $0.30 a comparison (derived) [18].

Expert work costs more. The GPQA authors (November 2023), who hired 61 contractors with or pursuing PhDs to write and validate graduate-level questions, estimate an average payment of about $95 an hour [19]. Humanity's Last Exam (January 2025) offered a $500,000 prize pool for a benchmark of 2,500 questions [20].

Amazon Mechanical Turk, the crowd marketplace two of the studies here used for their human baseline [18, 23], "permanently closed on September 30, 2026" [14]. This article infers that cost comparisons priced against Mechanical Turk lost their live baseline on that date. No source reviewed measures label prices since.

Paying for labels now includes checking them. One 2023 study estimated that 33% to 46% of crowd workers on a summarization task used LLMs [21]. Another, in 2021, estimated an average of at least 3.3% label errors in the test sets of 10 widely used datasets [22]. Agreement between labelers is one check, covered in measuring label quality with inter-annotator agreement.

Model-made labels and their measured cost

Labels from a model cost a fraction of human labels at the model prices of each study's year. Gilardi et al. (2023) found that a commercial chat model family's per-annotation cost was "less than $0.003", "about thirty times cheaper than MTurk", with zero-shot accuracy about 25 percentage points above crowd workers on average [23]. Wang et al. found it cost "50% to 96% less" to reach the same downstream performance with model labels in 2021, counting only the per-label charge [24].

Cheap labels do not always match a full human set. In Ding et al., 6,000 model labels costing $22.63 beat 3,000 human labels estimated at $221 to $300 on accuracy, while the full set of 67,349 human labels scored higher than both [25].

Generated data can be cheap in cash and costly in rights. Stanford's Alpaca team, in March 2023, generated 52,000 instructions and outputs for less than $500 through a commercial model's API, then barred commercial use, in part because that model's terms of use prohibit building competing models [26]. A 2024 Nature study found that indiscriminate training on model-generated content "causes irreversible defects in the resulting models" [27]. Model labels come from inference, through an API or a self-hosted model, and the break-even between the two is worked through in self-hosted LLM vs API cost.

Free data still costs compute and checking

Common Crawl's site, viewed on 8 November 2026, describes "a free, open repository of web crawl data" of over 300 billion pages [10]. Preparing it is the expense. The FineWeb team (2024) trained over 70 models, an estimated 80,000 H100 GPU hours, to choose the filters for a 15-trillion-token dataset from 96 Common Crawl snapshots, and spent 6,000 more H100 GPU hours scoring it for educational quality [12].

Rights checking costs too. A 2023 audit of more than 1,800 text datasets found license omission of 70% or more and error rates of 50% or more on widely used hosting sites [35].

The cost after delivery

Licenses end. Reddit's remaining performance obligations in contracts longer than a year, mostly long-term content licensing contracts, fell from $143.7 million at the end of 2025 to $62.4 million at 30 September 2026, and its 10-Q says the agreements with the two partners behind substantially all of its licensing contract value "are approaching the end of their respective current terms" [2, 3].

Evaluation repeats. AlpacaFarm (2023) reports a recurring $242 to evaluate each model on 805 instructions with human annotators [18]. Maintenance adds up: Sculley et al. wrote in 2015 that "it is common to incur massive ongoing maintenance costs in real-world ML systems" [30].

Supply limits on public text

A projection published in June 2024 puts the point where training datasets reach the size of the stock of public human text between 2026 and 2032, with a median year of 2028 [33]. A 2024 audit of 14,000 web domains found that a single year of new restrictions left "~5%+ of all tokens in C4, or 28%+ of the most actively maintained, critical sources in C4, fully restricted from use" [34].

What sets the cost of a custom dataset

Across the evidence, cost follows scope. The sources set the license and its terms. The fields and labels set the labeling cost. History and refresh set how often data is bought, labeled and delivered again. A build also has to resolve the entities that appear across its sources, the work described in how entity resolution matches records.

A custom dataset from Rillor's data services is specified by the same items: the sources in scope, the entities and fields, the format, how far back the records reach, the regions, the refresh cadence and the delivery, and it comes with its schema, source provenance, lineage, version history, quality notes and stated limits. Writing that scope down before a build starts is the subject of how to specify a training dataset.

Read the full report

The full report adds each filing's wording, the evidence classes, the label cost tables and the limits of every figure.