How training data is priced

Training data is sold in four ways: per task or item, per hour of human time, as a fixed price per dataset, or as a license fee to use data that already exists. Labeling software adds a fifth, a seat or usage subscription. Published rates range from $0.015 per labeled object to hundreds of dollars per expert-built task.

Last verified 7 November 2026. Rates are each company's published starting prices, not quotes.

What does each pricing model look like?

Model	You pay per	Published examples	Suits	Watch for
Per task or item	Labeled object, entity, ranking, finished task	Label Your Data: keypoints from $0.015 per object, bounding boxes from $0.02, NLP from $0.02 per entity (Label Your Data pricing page)	Well-defined, repetitive labels	Vendor speeds up on hard items to protect margin
Per hour	Annotator, expert or participant time	Label Your Data: hourly projects typically $6 per annotator hour. Prolific: participants paid at least $8 an hour ($12 recommended) plus a platform fee (Label Your Data; Prolific help center)	Tasks whose length varies, expert review, research studies	Slow work costs you more; needs throughput targets
Platform usage or seat	Software you run with your own annotators	Segments.ai: Core plan from 3,600 hours a year, equivalent to $2.67 an hour. HumanSignal: twelve users on Starter Cloud for $99 monthly (Segments.ai and HumanSignal pricing pages)	Teams with in-house annotators	Labor is still yours to hire and manage
Per dataset (fixed price)	A defined deliverable at an agreed size and quality bar	Rarely published. RL tasks with graders: "$200 to $2000 mostly" each, from Epoch AI interviews (Epoch AI, Jan 2026)	Scoped projects with a clear acceptance test	Change requests are repriced; scope tightly
License fee	Permission to train on existing data, for a term	Video footage: $1 to $4 a minute (Bloomberg). Reddit: $203.0M aggregate contract value over two to three years (S-1) (Bloomberg; TechCrunch on Reddit's S-1)	Breadth that already exists	Term limits, exclusivity premiums, unclear sublicensing

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

Per task: when the unit is clear

Per-item pricing fits labels that look alike from one item to the next: boxes around vehicles, named entities in contracts, a yes or no on a model answer. Slow items are the vendor\'s problem, not yours. Ask how odd items get handled; per-item pay rewards rushing them.

Starting prices are minimums. Label Your Data's rates are published as "from" prices, so a real project can only come in at or above them.

Per hour: when the work varies

Hourly pricing suits expert review, writing tasks and research studies where one item takes two minutes and the next takes twenty. You carry the speed risk, so agree a throughput estimate in the pilot. On Prolific, the hourly rate goes to participants and the platform fee is added on top; the fee rate was not stated on the page we read.

Per dataset: when you can define done

A fixed price for a dataset works when you can write an acceptance test: a size, a schema, an agreement rate between annotators, a pass rate against a grader. Without one, every disagreement about quality becomes a disagreement about money.

RL environments show this best. Epoch AI's January 2026 FAQ drew on market interviews: most graded tasks were said to sell for $200 to $2,000, and a $20,000 task was described as unusual yet possible.

License fees: when the data already exists

License fees pay for permission rather than labor. Price depends on size, uniqueness, exclusivity and term, and public numbers are scarce: Bloomberg reported AI companies paying $1 to $4 a minute for unused video footage, and Reddit's S-1 disclosed $203.0 million in aggregate contract value from data licensing.

Marketplace prices are set by the providers who list there. Snowflake Marketplace lets providers charge by usage (per query, billable event or monthly fee), by subscription or in installments.

How to compare quotes on different models

Convert every quote to cost per accepted item: the price, plus the cost of items rejected at review, plus your own team's review time, divided by the items you keep. The cost estimator does the first part with published rates.

Sources behind this section

Pricing questions

Which pricing model is cheapest?

The one whose unit matches how your cost actually behaves. Per-item pricing is cheapest for uniform labels; per-hour is fairer when items vary a lot; a fixed price works only when you can write an acceptance test.

Why do so few vendors publish prices?

Quality bars, expert type, languages and security requirements change the cost of the same task several times over. Most vendors scope first and quote after, which is why the directory lists so many as "not published".

What does exclusivity add to the price?

In RL environments, sources quoted by Epoch AI said an exclusive deal costs around four or five times the shared equivalent. For licensed data, exclusivity is negotiated case by case; Labelbox Prism, for example, structures its licenses as exclusive by default.

Are platform fees on marketplaces paid by the buyer?

Usually the provider pays them out of the price you see. AWS Data Exchange charges providers a 3% listing fee on public offers; Datarade charges providers between 15% and 30% commission depending on plan. TollBit is an exception: it adds a small percentage on top for developers.

Should a pilot be paid?

Yes. A paid pilot of a few hundred items gets the vendor's real team and real process, and gives you a measured cost per accepted item to compare against the quote.