Computed 17 Oct 2026 from 1,022 observations
What AI training data goes for, on each side of the market.
The same hour of speech is paid to the person who records it, offered to the company that holds it, listed in a catalogue and sold to a lab, each at a different price. Here is the median at each of those points, per comparable unit, from every sourced figure we hold.
The market on one scale
Every kind of data, from a cent to a company
Every commercial public price in each kind's main unit, one mark each, on one money scale. The ink mark is the median and the band the middle half, computed the way the tables below are. Each row opens its kind.
Company recordsper company
$316K median of 97
Datasets, tasks and environmentsper environment
$2,000 median of 33
Source codeper repository
$1,579 median of 9
Text and archivesper book or title
$515 median of 18
Audio and speechper hour of audio
$118 median of 132
Expert and human workper hour of work
$77 median of 114
Video and footageper hour of video
$37 median of 45
Physical-world dataper hour of video
$16 median of 56
Imagesper image
$0.53 median of 46
The spread
Where the evidence holds a median of 8 or more figures at both ends of a like-for-like pair in the same unit: what the people who make data are paid against what buyers pay for it made to order, and what holders are offered against what the same kind of data lists or sells for. The gap is the room a direct deal has.
Datasets, tasks and environments, per task
9.5x
Paid to the people who make it: $300 median (11 figures). Made to order: $2,837 median (8 figures).
Physical-world data, per hour of video
2.7x
Paid to the people who make it: $14 median (14 figures). Made to order: $38 median (20 figures).
Work by people, per hour of work
0.2x
Paid to the people who make it: $77 median (105 figures). Made to order: $19 median (9 figures).
Company records
The record of how a business ran: documents, tickets, communications, sold or licensed as a whole.
Offered to holdersper company
$316K
Middle half $174K to $693K. 93 figures (+8 open-ended), 13 Aug 2026 to 17 Oct 2026.
The figures
- An operating company's or nonprofit's records (email, documents, chat, calls and transcripts, CRM, billing and operations), licensed to Mode (BizBuyData) for AI training $707K
- Operating company's records, Mode (BizBuyData) estimator output: 100 full-time desk employees, 0 non-desk, 10 years of operational data, team based in the EU, Canada or Australia $300K, eu canada australia
- Operating company's records, Mode (BizBuyData) estimator output: 100 full-time desk employees, 0 non-desk, 10 years of operational data, team based outside the USA, EU, Canada and Australia $124K, other
- Operating company's records, Mode (BizBuyData) estimator output: 100 full-time desk employees, 0 non-desk, 10 years of operational data, team based in the USA $407K, us
- Operating company's records, Mode (BizBuyData) estimator output: 20 full-time desk employees, 0 non-desk, 3 years of operational data, team based in the USA $187K, us
- Operating company's records, Mode (BizBuyData) estimator output: 50 full-time desk employees, 0 non-desk, 10 years of operational data, team based in the USA $311K, us
- One healthcare company's operating records (clinical workflows, billing appeals and other records it already kept), licensed through Shfa.ai to AI buyers $943K, healthcare
- A medical facility's data (records it keeps in operation), licensed for AI training (Shfa.ai's headline range) $949K, healthcare
- Operating company's records, Pewter estimator output: 100 employees, 10 years in business, US firm (Pewter's licensing partner's payout model) $587K, 100 employees
- A company's operational and business data licensed to Ooak Data (headline) Earn up to $1M
- A company's workspace data (email, chat, meetings, CRM), anonymised and priced by Corpus $447K
- Largest company workspace-data licence Corpus has closed $2M
- Operational data from a company's systems (Salesforce, Jira, SAP, ServiceNow) licensed through Labelbox Prism $224K
- A general contractor's operational record, perpetual exclusive licence (Polyshares) $300K, construction
- A global marketing agency's operational record, perpetual exclusive licence (Polyshares) $110K, marketing agency
- A roofing company's operational record, perpetual exclusive licence (Polyshares) $450K, home services
- A solar company's operational record, perpetual exclusive licence (Polyshares) $275K, energy
- A company's operating record (Slack, email, tickets, CRM, files), bought by Polyshares $447K
- A third-party logistics firm's dispatch, customer messages, SOPs, logistics workflows and operating conversations (Red Deer Data, completed deal) $500K, logistics
- An accounting group's workpapers, client service workflows, templates, internal communications and operating documents (Red Deer Data, completed deal) $280K, accounting
- A company's business software records (communications, orders, finance, SOPs), privately bid by Red Deer Data $224K
- An industrial distributor's communications, product knowledge, order history, supplier records and operating documents (Red Deer Data, completed deal) $1.2M, distribution
- A company's operating data (code, docs, CRM, finance, support history), company of 20 to 50 employees (Replay) $31.6K, 20 50 employees
- A company's operating data (code, docs, CRM, finance, support history), company of 250 or more employees (Replay) $1,000,000+, 250 plus employees
- A company's operating data (code, docs, CRM, finance, support history), company of 50 to 250 employees (Replay) $316K, 50 250 employees
- A US business's documents, conversations and operational records (Telegraph Lab) $632K
- Offers to small operating businesses for their internal records (email, Slack, documents), from unnamed AI-data buyers and brokers six-figure "exploding offers" that expire in 48 hours, operating
- The same closed crypto startup's workspace data (Google, Notion, Figma and Slack), sale through SimpleClosure pending $15K, workspace
- Illustrative valuation of a logistics company's systems (2.1M emails, 1.4M Slack messages, 96k tickets, 86k files, 64k invoices, 12k workbooks) by the broker Corpus Bridge $153K
- A company's data (chat logs, codebases, docs, ledgers) listed with the Grepped broker $316K
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 10 employees, founded 2020, $0 to $1M revenue $16.9K, 10 employees
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 100 employees, founded 2016, $10M to $50M revenue $85.6K, 100 employees 10 50m revenue
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 100 employees, founded 2016, revenue not disclosed $45K, 100 employees
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 1000 employees, founded 2001, $200M+ revenue $468K, 1000 employees
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 200 employees, founded 2016, revenue not disclosed $49.7K, 200 employees
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 25 employees, founded 2016, $1M to $10M revenue $34.7K, 25 employees
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 250 employees, founded 2011, $50M to $200M revenue $222K, 250 employees
- A company's data (chat, code, docs, ledgers), Grepped owner estimator output: 50 employees, founded 2016, $1M to $10M revenue $40.1K, 50 employees
- A company's data, Grepped referrer calculator output: Accounting & bookkeeping, team of 30, 10 years in operation $191K, accounting
- A company's data, Grepped referrer calculator output: Healthcare, team of 100, 10 years in operation $693K, healthcare
- A company's data, Grepped referrer calculator output: Marketing & advertising agency, team of 20, 5 years in operation $58.3K, marketing agency
- A company's data, Grepped referrer calculator output: Semiconductors, team of 200, 10 years in operation $1.4M, semiconductors
- A company's data, Grepped referrer calculator output: Software & SaaS, team of 100, 10 years in operation $574K, software 100 employees
- A company's data, Grepped referrer calculator output: Software & SaaS, team of 25, 5 years in operation $98.4K, software 25 employees
- Operating company's records, Handshake AI estimator output: 100 employees, 10 years of operation, Europe $440K, europe
- Operating company's records, Handshake AI estimator output: 100 employees, 10 years of operation, Other $176K, other
- Operating company's records, Handshake AI estimator output: 100 employees, 10 years of operation, USA $586K, 100 employees
- Operating company's records, Handshake AI estimator output: 150 employees, 15 years of operation, USA $882K, 150 employees
- Operating company's records, Handshake AI estimator output: 20 employees, 3 years of operation, country Other Up to $100,000, other small
- Operating company's records, Handshake AI estimator output: 20 employees, 3 years of operation, USA $172K, 20 employees
- Operating company's records, Handshake AI estimator output: 200+ employees, 20+ years of operation, USA $1.2M, 200 employees
- Operating company's records, Handshake AI estimator output: 30 employees, 5 years of operation, USA $246K, 30 employees
- Operating company's records, Handshake AI estimator output: 50 employees, 10 years of operation, USA $420K, 50 employees
- Operating company's records, Handshake AI estimator output: 50 employees, 5 years of operation, USA $336K, 50 employees 5 years
- An operating company's operational data (Handshake AI) $632K
- Operating company's records, micro1 estimator output: 100 employees, 10 years of data, headquarters Europe $435K, europe
- Operating company's records, micro1 estimator output: 100 employees, 10 years of data, headquarters Other $174K, other
- Operating company's records, micro1 estimator output: 100 employees, 10 years of data, headquarters US $580K, 100 employees
- Operating company's records, micro1 estimator output: 1000 employees, 30 years of data, headquarters US $5.4M, 1000 employees
- Operating company's records, micro1 estimator output: 20 employees, 3 years of data, headquarters US $222K, 20 employees
- Operating company's records, micro1 estimator output: 250 employees, 15 years of data, headquarters US $1.2M, 250 employees
- Operating company's records, micro1 estimator output: 30 employees, 5 years of data, headquarters US $277K, 30 employees
- Operating company's records, micro1 estimator output: 50 employees, 10 years of data, headquarters US $416K, 50 employees
- Operating company's records, micro1 estimator output: 50 employees, 5 years of data, headquarters US $333K, 50 employees 5 years
- Operating company's records, micro1 estimator output: 500 employees, 20 years of data, headquarters US $2.4M, 500 employees
- An operating company's operational data package (micro1 Enterprise Data Partnership, as stated on its referral page) $447K
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 10 to 24 FTEs, 3 to 7 years in business, headquarters United States $225K, 10 24 employees
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 101 to 250 FTEs, 15+ years in business, headquarters United States $762K, 101 250 employees
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 25 to 49 FTEs, 3 to 7 years in business, headquarters United States $275K, 25 49 employees 3 7 years
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 25 to 49 FTEs, 8 to 15 years in business, headquarters United States $341K, 25 49 employees
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 250+ FTEs, 15+ years in business, headquarters United States $1.3M, 250 plus employees
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 50 to 100 FTEs, 8 to 15 years in business, headquarters Canada, UK or Europe $285K, canada uk europe
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 50 to 100 FTEs, 8 to 15 years in business, headquarters International $94K, international
- Operating company's accounting, customer and scheduling records, Nyne estimator output: 50 to 100 FTEs, 8 to 15 years in business, headquarters United States $447K, 50 100 employees
- An established US business's accounting and customer records (Nyne) Commonly $100K or more
- A company's operational data, Ooak estimator output: 2+ years of history, 1-10 people at peak, Winding down, tools: Google Drive, Slack $9,381, 1 10 employees
- A company's operational data, Ooak estimator output: 7+ years of history, 101-200 people at peak, Operating, tools: Google Drive, Gmail, Slack, Jira, GitHub, HubSpot $206K, 101 200 employees
- A company's operational data, Ooak estimator output: 4+ years of history, 11-50 people at peak, Closed, tools: Google Drive, Gmail, Slack, Jira, GitHub, HubSpot $39.4K, closed
- A company's operational data, Ooak estimator output: 4+ years of history, 11-50 people at peak, Operating, tools: Google Drive, Gmail, Slack, Jira, GitHub, HubSpot $60.2K, 11 50 employees
- A company's operational data, Ooak estimator output: 7+ years of history, 200+ people at peak, Operating, tools: Google Drive, Gmail, Slack, Jira, GitHub, HubSpot $273K, 200 plus employees
- A company's operational data, Ooak estimator output: 4+ years of history, 51-100 people at peak, Operating, tools: Jira, GitHub, HubSpot $18.8K, no docs no email
- A company's operational data, Ooak estimator output: 4+ years of history, 51-100 people at peak, Operating, tools: Google Drive, Gmail, Slack, Jira, GitHub, HubSpot $101K, 51 100 employees 4 years
- A company's operational data, Ooak estimator output: 7+ years of history, 51-100 people at peak, Operating, tools: Google Drive, Gmail, Slack, Jira, GitHub, HubSpot $117K, 51 100 employees
- Structured, de-identified company records (Scale AI) $100K
- A business's operational records licensed through SuperAnnotate Data Partnerships $316K
- A company's operational data, Troveo estimator output: 100 employees, 8 to 15 years of operating $894K, 100 employees
- A company's operational data, Troveo estimator output: 1,000+ employees, 15+ years of operating $3.3M, 1000 plus employees
- A company's operational data, Troveo estimator output: 20 employees, 3 to 7 years of operating $346K, 20 employees
- A company's operational data, Troveo estimator output: 250 employees, 8 to 15 years of operating $1.3M, 250 employees
- A company's operational data, Troveo estimator output: 500 employees, 8 to 15 years of operating $1.8M, 500 employees
- A company's whole working record (Turing) Up to $1M
- An operating company's records (micro1) $316K
- A bid Grepped says it received for a small company's records: an ~8-person team with email, chat, code, calls and docs over 3 years $50K, 8 employees
- Spirit Airlines' corporate records archive, micro1's late competing proposal after the auction (in-house de-identification) $12.5M, bankrupt
- A full company's operational data licensed through Troveo's business data program Typical deals start at six figures
- Spirit Airlines' de-identified internal records (Microsoft 365 email, Teams, OneDrive, SharePoint, employee records), Google's opening auction bid $5M
- Spirit Airlines' de-identified internal records, Mercor's later bid conditioned on in-house de-identification $10M
- Spirit Airlines' de-identified enterprise data, Mercor's alternate bid under its signed alternate sale agreement $7.5M, bankrupt
- Spirit Airlines' de-identified internal records, Mercor's first-round alternative conditioned on de-identifying with its own tools $7M
- Spirit Airlines' de-identified internal records, Mercor's first overbid on Google's terms (third-party de-identification) $5.2M
- A startup's code and communications after its team was hired away (Warmly) Four offers, up to $300K
Offered to holdersper document or page
$50too few figures for a median
1 figure, 24 Mar 2026.
Offered to holdersper contract
Open-ended only
0 figures (+1 open-ended), 3 Sep 2026.
The figures
- Proceeds to venture-backed startups from AssetHub sales and licences of code and workspace data (aggregate, 350+ companies supported) more than $10 million (on track, by end of 2026)
Sales that happenedper company
$28.3K
Middle half $21.3K to $2.5M. 4 figures (+4 open-ended), Apr 2026 to 14 Oct 2026.
The figures
- Company data licensed through Ooak from $20K to several hundred thousand dollars
- A closed crypto startup's code base (Keelan Miskell's company, about two and a half years of work), sold through SimpleClosure's Asset Hub $25K, code
- Company assets (code and workspace data) sold or licensed through SimpleClosure's AssetHub deals in the tens to hundreds of thousands of dollars
- Spirit Airlines' de-identified enterprise data (email, Teams, OneDrive, SharePoint, CRM, revenue management, workflow data), Google's successful bid under the signed sale agreement $10M, bankrupt
- A shut-down startup's archive (SimpleClosure's first deals) $31.6K, closed
- Code bases and workspace data of shut-down startups sold through SimpleClosure's test marketplace, the average deal over its first year $10K, closed
- Defunct companies' workplace data (Slack, email, Jira, documents) bought by the wind-down firm Sunset a similar price range (同様の価格帯) to SimpleClosure's 1万ドル to 10万ドル ($10,000 to $100,000) per company, closed
- Data assets of a closed transcription company built over more than a decade (cielo24, est. 2012; per Forbes, 13 years of Slack messages, Jira tickets, emails and Google Drive files), sold during its wind-down to an organisation in language technology mid-six figures, closed
By segment, where two or more figures exist
- Bankruptoffered to holders, per company$7.5M to $12.5M2 figures
- 1000 employeesoffered to holders, per company$468K to $5.4M2 figures
- 500 employeesoffered to holders, per company$1.8M to $2.4M2 figures
- 250 employeesoffered to holders, per company$1.2M3 figures
- 200 employeesoffered to holders, per company$49.7K to $1.2M2 figures
- Healthcareoffered to holders, per company$943K to $949K2 figures
- 100 employeesoffered to holders, per company$586K5 figures
- Europeoffered to holders, per company$435K to $440K2 figures
- 50 employeesoffered to holders, per company$416K3 figures
- 50 employees 5 yearsoffered to holders, per company$333K to $336K2 figures
- USoffered to holders, per company$311K3 figures
- Accountingoffered to holders, per company$191K to $280K2 figures
- 30 employeesoffered to holders, per company$246K to $277K2 figures
- 20 employeesoffered to holders, per company$222K3 figures
- Otheroffered to holders, per company$174K3 figures
- Marketing agencyoffered to holders, per company$58.3K to $110K2 figures
- Closedsales that happened, per company$10K to $31.6K2 figures
Source code
Repositories with their history. Common code is cheap; scarce languages and linked tickets are not.
Paid to the people who make itper task
$130
Middle half $99 to $730. 7 figures (+1 open-ended), 26 May 2026 to 15 Oct 2026.
The figures
- Review and upgrade of one real-world coding, data or automation scenario used to train AI assistants (Python, JavaScript, SQL, n8n, Make, Zapier), Vetto Coding, Data and Automations Experts $30, code review
- Original SWE-bench style coding task authored from the contributor's own private codebase (maintainer-style issue, hidden fail-to-pass and pass-to-pass tests, reference solution), AfterQuery University SWE Expert $75, software engineering task
- Reproducible coding RL environment (feature implementation or bug fix with a deterministic verifier and a golden reference solution), micro1 Software Engineer (GitHub Contributor) $1,061, coding rl environment
- One self-contained database engineering task for AI agents (engine implementation, production diagnosis or application migration) with a reproducible repository, reference solution, tests and grading criteria, Mercor Database Systems Evaluation Expert pilot $2,200, software engineering task
- Codebase-exploration RL task on a production Python repository (Django, Twisted, Scrapy, SymPy, mypy, pip and others), Mercor Codebase Q&A $130
- Frontier Code Agents project for an unnamed AI research lab: use AI coding agents to complete and evaluate realistic engineering tasks, review model-generated code for bugs and edge cases (Mercor, Backend Engineer, Coding Agent Experience) $400, coding
- One prompt-and-code-review task (writing a prompt, A/B comparing two models' code and picking the better) for Claude Code fine-tuning, Snorkel AI's project Marlin for Anthropic up to $280 per task, code review
- Code repository seeded with an OSS-fuzz-style vulnerability plus a validated patch, AfterQuery Cybersecurity Expert $122, security patch
Paid to the people who make itper hour of work
$118
Middle half $92 to $151. 4 figures, 26 Nov 2025 to 15 Sep 2026.
The figures
- Software engineering expert pool (writing engineering problems with reference answers, grading model-written code) $102, coding
- Application user recording real work in Visual Studio Code on macOS (computer-use data) $60, computer use
- Legacy codebase migration expert designing and grading migration tasks for a frontier AI lab $200, coding
- Coding-benchmark task work in React, TypeScript or JavaScript: curate issues, solutions and test suites from real-world repositories for AI labs (Mercor, Frontend Software Engineer), hourly base plus a bonus per approved task $134, coding
Offered to holdersper repository
$1,478
Middle half $934 to $24.2K. 6 figures (+2 open-ended), 24 Mar 2026 to 14 Oct 2026.
The figures
- Access to a substantial private codebase the contributor owns, used as the basis for SWE coding tasks (AfterQuery University SWE Expert) $300, software engineering task
- A well-documented, production-grade private codebase licensed on HUD's DataVendor marketplace up to $20,000 each
- A private repository of deep depth, licensed non-exclusively through Pangea's code licensing partner $1,732, deep
- A private repository of light depth, licensed non-exclusively through Pangea's code licensing partner $837, light
- A private repository of typical depth, licensed non-exclusively through Pangea's code licensing partner $1,225, typical
- A private repository bought by Turing's Project Lazarus for AI training (all languages, priced on quality and complexity) $31.6K, coding
- A private repository, graded by the platform $5,000 baseline
- Legacy private codebase (COBOL, .NET, Java) licensed to Turing for AI training $31.6K, legacy
Offered to holdersper token
$0.00004 to $0.0001too few figures for a median
2 figures, 14 Oct 2026.
The figures
Offered to holdersper pull request
$1,000too few figures for a median
1 figure, 24 Mar 2026.
The figures
Offered to holdersper contract year
Open-ended only
0 figures (+1 open-ended), 1 Oct 2025.
The figures
- Professional developers' code snippets, edit history, prompts and terminal usage shared with JetBrains for AI training free All Products Pack subscriptions for one year ... currently priced at $979.00 per user/year
Sales that happenedper repository
$1,579too few figures for a median
1 figure (+1 open-ended), 14 Sep 2026 to 14 Oct 2026.
The figures
- Private codebases licensed through Pangea's network to date: total paid to owners $1,579
- A failed startup's codebase Tens of thousands of dollars, closed
Asking prices off the shelfper dataset
$150
Middle half $125 to $98.1K. 3 figures (+1 open-ended), 21 Sep 2026 to 11 Oct 2026.
The figures
- AtCoder competitive-programming dataset (problems, test data, accepted and failed user solutions) licensed to AI companies budget is $100,000 or more, competitive programming
- Legacy software codebases across 20+ industries (400+ codebases, 204M+ lines, 420K+ commits in the full set), price for an initial batch of 20 million lines $196K, coding
- Code Agent Trajectories: 551 verified PASS trajectories from an autonomous code agent in a sandboxed Python environment (reasoning, tool calls, final answer; JSONL, 5.63 MB), sold by Jurgen's Data Lab on Getly $100, agent trajectory
- Premium: 177 verified PASS agent trajectories that also carry full chain-of-thought reasoning (JSONL, ATIF-v0.2, 1.3 MB), sold by Jurgen's Data Lab on Getly $150, agent trajectory
Asking prices off the shelfper repository
$99 to $65.8Ktoo few figures for a median
2 figures (+1 open-ended), 2 Sep 2026 to 11 Oct 2026.
The figures
- Data-structures-and-algorithms code repositories with programming solutions, price for an initial batch of 1,000 repositories $99, coding
- 240+ private commercial repositories with full git history (300,000+ commits, ~42M lines, 44,000+ merged MRs with descriptions, ~463M tokens), de-identified $263,200, coding
- Data-Analysis-Platform: one private multi-service enterprise Git repository with full history (5,683 commits, 1,474 tracked files, about 108,761 lines; Terraform, Python backend, IAM, SQL migrations, API tests), sold by Octobit8 on Opendatabay $65.8K, git history
Asking prices off the shelfper record
$0.0025too few figures for a median
1 figure, 11 Oct 2026.
Made to orderper record
$8.94too few figures for a median
1 figure, 14 Oct 2026.
Made to orderper dataset
Open-ended only
0 figures (+1 open-ended), 14 Oct 2026.
Cost to produceper record
$0.5too few figures for a median
1 figure, 25 Feb 2026.
The figures
By segment, where two or more figures exist
- Software engineering taskpaid to the people who make it, per task$75 to $2,2002 figures
- Codingpaid to the people who make it, per hour of work$1343 figures
- Agent trajectoryasking prices off the shelf, per dataset$100 to $1502 figures
Video and footage
Footage licensed by the minute or the hour. Rights and specificity set the price, not volume.
Paid to the people who make itper hour of video
$2.97
Middle half $2.85 to $3.75. 4 figures, 11 Jun 2026 to 9 Oct 2026.
The figures
- First-person household-task video recorded by a gig worker in one week through the Atlas Capture app (Mecka AI), paid in a US-dollar stablecoin $2.94, egocentric
- First-person household-chore video recorded on a smartphone through the Human Motion Data Recorder app (contributors largely in the Philippines), per approved hour $3.00, egocentric
- LATAM Egocentric Project: first-person video of everyday tasks recorded by contributors in Latin America, up to 100 hours each $6.00, egocentric
- First-person (head-mounted phone) video of household tasks recorded at home in Chennai, India, for robot-training data buyers $2.60, egocentric
Paid to the people who make itper task
$18
Middle half $11 to $84. 3 figures, Aug 2026 to 11 Oct 2026.
The figures
- Natural conversation video recording study (two participants on video call), United States $18, faces
- In-lab movement recording study (camera video, motion sensors, mmWave radar, audio), Las Vegas $150, human motion
- Selfie video collection, 4 videos (2 indoor, 2 outdoor) with head movement and facial expressions, Malaysia, Singapore, Hong Kong $3.01, faces
Paid to the people who make itper hour of work
$12 to $17too few figures for a median
2 figures, 3 Sep 2026 to 4 Sep 2026.
The figures
Offered to holdersper hour of video
$95
Middle half $87 to $133. 7 figures, 11 Apr 2024 to 16 Oct 2026.
The figures
- Raw, unlabelled video recordings sold to AI developers through the FileYield seller marketplace (bulk material for generative pre-training) $95, raw video
- User-generated news and viral video archives uploaded to Newsflare for licensing to AI companies' training sets $22, news
- Creator video licensing rates, ordinary to professional or niche content $147, creator
- Creators' unused footage licensed through Troveo for AI training (Troveo's typical range) $90, creator
- Most unused creator footage made for YouTube, Instagram and TikTok $85, creator
- A creator's unused footage $120, creator
- Short clips of people's everyday actions, emotions and body parts requested by Adobe for its video model (top of its pay) $435, human action
Offered to holdersper task
$120too few figures for a median
1 figure, 11 Apr 2024.
The figures
Offered to holdersper record
Open-ended only
0 figures (+1 open-ended), 5 Apr 2024.
The figures
- Videos from Photobucket's 13 billion photo and video archive, in licensing talks more than $1 per video, photo host
Sales that happenedper hour of video
$42
Middle half $14 to $113. 6 figures, 11 Apr 2024 to 14 Oct 2026.
The figures
- Troveo's cumulative payouts to video content owners for AI-training licences, against the hours of video it has licensed (creator, studio and archive footage) $7.14, creator
- Premium, metadata-rich film, TV and archive content licensed to AI platforms (MIPCOM 2025 benchmark) $134, film
- A YouTuber's decade-old music videos licensed to AI companies (about 1,000 hours in a year) $33, creator
- Everything Troveo paid content owners in its first year: $5 million for 1 million hours of footage $5.00, creator
- One creator's back catalogue monetised through Troveo (900 hours of video) $51, creator
- Video footage bought by Adobe for a text-to-video model $157, creator
Sales that happenedper contract
Open-ended only
0 figures (+4 open-ended), 8 Oct 2025 to 26 Jun 2026.
The figures
- Crowd-sourced photo, video and audio dataset licence from the Kled AI marketplace $12M non-exclusive data licensing deal over two years
- Terabytes of creator video licensed through Troveo, deals negotiated by a creator manager (Night Media) about five deals, each worth tens of thousands of dollars, creator
- A large creator's video catalogue licensed through Troveo, total paid to each of a handful of creators more than $1 million each, creator
- An EMEA broadcaster's scripted and cinematic library licensed for AI training through Protege over $1M in 6 months, film
Sales that happenedper contract year
Open-ended only
0 figures (+1 open-ended), 17 Oct 2025.
The figures
- A regional sports rights distributor's archival footage licensed to generative AI companies through Protege $250,000+ in year one, sports
Asking prices off the shelfper hour of video
$37
Middle half $15 to $99. 17 figures (+5 open-ended), 5 Apr 2024 to 14 Oct 2026.
The figures
- Stockfilm archival footage dataset (162,072 digitized clips, 1930s to 1980s amateur film, 298+ hours, 121 countries) with structured metadata, commercial AI training licence, volume-tiered From $0.95 per clip, stock
- Unidata DeepFake Videos Dataset, 5,000 real videos with AI-generated faces, one per person, mean 9 s Starts at $5,000 / purchase, faces
- Unidata Fabric Masks and Disguise Presentation Attack Dataset, 9,270 videos of 103 consented people, 3 to 5 s, 1080p to 4K Starts at $30,000 / purchase, 4k
- Unidata iBeta Level 1 presentation-attack dataset, 35,800 videos of 50 actors, ~5 s each, up to 4K Starts at $30,000 / purchase, faces
- DrivingFootage.com POV Cycling, 1,710 1-minute stabilized 4K clips from 34 US locations, 28+ hours $85, 4k
- DrivingFootage.com Colorado & Wyoming, 6,516 2-minute 4K 360-degree POV driving clips, 215+ hours $101, driving
- DrivingFootage.com Los Angeles, 9,114 2-minute 4K 360-degree POV driving clips, 310+ hours $99, driving
- DrivingFootage.com Full Library, 50,000 4K 360-degree POV driving clips, 1,750+ hours, 118 TB $86, driving
- AKYS Pathfinder georeferenced LWIR thermal driving video, Lithuania, 82 minutes $190, driving
- Shofo RecipeBook self-serve video training data, semantic search over 25M+ clips, bought by the hour $3.00, web video
- Infobay.AI 2.2K+ Hours English and Japanese Vertical Storyline Video Dataset (scripted and natural short-drama style) $30, vertical drama
- Infobay.AI 106K+ Hours STEM Classroom Video Dataset (price for initial 10,000-hour batch) $15, classroom
- Infobay.AI 30K+ Hours Vertical User-Generated Content (UGC) Video Dataset, smartphone vertical videos across lifestyle, travel, food and more $15, user-made
- Wreck Room Wreck-7K, ~7,000 hours of 1080p exocentric human-object interaction video (object destruction sessions, Las Vegas) $10, human action
- Princep Up to 20,000 Hours of real English online job-interview video (global accents) with transcripts $2.99, interview
- Princep 8,365 Hours Indian English real online job-interview video with transcripts, 31,826 consented speakers $2.83, interview
- Rights-cleared archive video (sports federations', broadcasters' and other rights holders' footage) licensed to AI developers through the Veritone Data Refinery, priced per hour by application and metadata $32, broadcaster
- JoinAI Multilingual Clean Talking-Head Video Dataset, 841 synthetic 1080p clips, ~3.5 hours $37, synthetic
- UltraSports4K, 5,200 professionally produced 4K sports/action videos, 15 to 20 s each, ~737 GB $632, stock footage
- THVD Talking Head Full Video Dataset, 47,547 videos, 500+ hours, 20,841 identities, 60% 4K $72, faces
- Film and TV content licensed through Calliope Networks (now part of Protege) $6.25 per minute for HD content, more for 4K or 3D, film
- Longer films and professional video (Defined.ai's quoted rate) $173, film
Asking prices off the shelfper record
$2.50 to $2.83too few figures for a median
2 figures (+2 open-ended), 6 Apr 2024 to 11 Oct 2026.
The figures
- Unidata Anti-Spoofing Replay Dataset, 42,743 video clips (phone and PC recordings) for liveness detection Starts at $10,000 / purchase, faces
- Unidata Phone and Webcam Video Dataset, 30,952 videos and photos of 3,869 people for anti-spoofing, ~30 s and ~8 s takes Starts at $10,000 / purchase, faces
- CapitalPower Aerial Video of Nigeria Cities, ~500 4K drone video files of urban landscapes with GPS metadata $2.50, drone
- Short-form videos licensed for AI training (Defined.ai) $2.83, short clip
Asking prices off the shelfper dataset
$987Ktoo few figures for a median
1 figure, 14 Aug 2026.
Asking prices off the shelf (research use only)per hour of video
Open-ended only
0 figures (+1 open-ended), 14 Oct 2026.
Made to orderper hour of video
$32
Middle half $23 to $95. 11 figures, 27 Mar 2026 to 22 Sep 2026.
The figures
- Crowdsourced video data capture (planning figure) $80, crowdsourced
- Studio or onsite video data capture (planning figure) $150, studio
- Made-to-order egocentric video, finished hour, Aria-class research glasses (multi-camera RGB, SLAM, eye and hand tracking) $32, egocentric
- Made-to-order egocentric video, finished hour, GoPro head or chest mount mono RGB rig $17, egocentric
- Made-to-order egocentric video, finished hour, Meta Quest 3 headset capture (stereo passthrough, IMU, hand tracking) $22, egocentric
- Made-to-order egocentric video, finished hour, Intel RealSense D435i/D455 stereo depth helmet rig $24, egocentric
- Made-to-order paired egocentric plus exocentric human video capture (moving head camera synchronised with fixed external cameras), per capture hour $38, ego exo
- Video object tracking annotation project, 100 hours, 5 objects tracked $110, annotation
- Crowd-platform video collection for simple tasks, without domain QA $11, crowdsourced
- Managed collection of general video (dashcam, ambient scenes, consumer camera) with standard QA $24, general video
- Annotation of egocentric and manipulation footage (action segmentation, object boxes, task-completion labels) $509, annotation
By segment, where two or more figures exist
- Egocentricpaid to the people who make it, per hour of video$2.974 figures
- Facespaid to the people who make it, per task$3.01 to $182 figures
- Creatoroffered to holders, per hour of video$1054 figures
- Creatorsales that happened, per hour of video$335 figures
- Drivingasking prices off the shelf, per hour of video$1004 figures
- Interviewasking prices off the shelf, per hour of video$2.83 to $2.992 figures
- Annotationmade to order, per hour of video$110 to $5092 figures
- Crowdsourcedmade to order, per hour of video$11 to $802 figures
- Egocentricmade to order, per hour of video$234 figures
Audio and speech
Recorded speech by the hour. Conversations, consent and uncommon languages are what buyers pay for.
Paid to the people who make itper hour of work
$24
Middle half $14 to $42. 31 figures, 4 Dec 2023 to 14 Oct 2026.
The figures
- Median advertised hourly pay for voice acting and voice-over roles in AI training $35, voice acting
- AI Tutor specialised in multilingual audio, Yoruba (xAI): records, annotates and transcribes Yoruba speech and localises English text into Yoruba $40, yoruba
- British English recordings evaluated for pronunciation, fluency, tone and audio quality $51, english
- Spanish (Mexico) voice actor recording scripted and emotional audio for AI training $36, spanish
- Bengali audio and video transcription and machine-transcript review $15, bengali
- Gujarati audio transcription and review for AI training $9.80, gujarati
- Norwegian audio and video transcription and speech-segment classification $24, norwegian
- Punjabi audio and video transcription with speaker identification and metadata $11, punjabi
- Expressive voice recording for AI training, Bengali speakers in India (Outlier, via Braintrust) $7.50, bengali
- Expressive voice recording for AI training, Simplified Chinese speakers (Outlier, via Braintrust) $20, mandarin
- Expressive voice recording for AI training, French speakers (Outlier, via Braintrust) $25, french
- Expressive voice recording for AI training, Thai speakers (Outlier, via Braintrust) $10, thai
- Expressive voice recording for AI training, Vietnamese speakers in Vietnam (Outlier, via Braintrust) $11, vietnamese
- Japanese audio transcription, error correction and context notes for AI training $27, japanese
- English phone calls to US businesses (reservations, appointments) made by Latin American contractors to help train AI systems $22, english
- Mandarin and English audio transcription and review for AI training $24, mandarin
- Portuguese audio transcription and review, with occasional English $13, portuguese
- Bilingual Hebrew and English audio contributor for AI training (AfterQuery) $30, hebrew
- Paid English conversation recording study for AI training (RemoExperts, open worldwide) $20, english
- Tamil audio transcription and review, varied accents and dialects $15, tamil
- Mandarin-English code-switched conversation recording (1 to 2 minute clips) by speakers in Singapore or Malaysia $3.16, mandarin english
- German speech data: recording audio samples, phonetic annotation and evaluation $44, german
- Voice actor recording a customer-service agent voice for cloning, African American English (Mercor) $71, english
- Voice actor recording a customer-service agent voice for cloning, Swiss German (Mercor) $87, swiss german
- Voice recorder, German and English (micro1) $44, german
- Audio recording contributor with an Australian English accent (AfterQuery) $50, english
- French conversation partner recording natural conversations for AI training (Babel Audio, Canada and US) $18, french
- Hindi voice actor or actress recording for AI training (Babel Audio) $95, hindi
- Conversation recording roles on David AI's Babel Audio platform $26, voice
- Voice actor recording a customer-service agent voice for cloning, US English (Mercor) $87, english
- Rural Indian speakers recording speech data for Karya (supplier to Microsoft and Google) $5.00, voice
Paid to the people who make itper hour of audio
$30
Middle half $21 to $115. 3 figures, 26 Sep 2025 to 14 Oct 2026.
The figures
- NaijaVoices: the rate it costs per hour of recording by its Nigerian 'data farmers' (Hausa, Igbo, Yoruba read speech) $30, african
- Spanish voice-over recording for AI training and media, by finished hour of audio (professional home studio) $200, spanish
- Recorded phone-call audio sold to AI developers (Neon Mobile app, paid to callers) $13, phone calls
Paid to the people who make itper contract
Open-ended only
0 figures (+1 open-ended), 22 May 2026.
Sales that happenedper hour of audio
$76
Middle half $40 to $2,005. 9 figures, Dec 2018 to 6 Oct 2026.
The figures
- Voice owner's royalty when others generate speech with their cloned voice in the ElevenLabs Voice Library (default rate) $1.20, voice licensing
- Voice owner's royalty in the ElevenLabs Voice Library for HQ or rare voices (maximum requestable rate) $8.00, voice licensing
- One hour of Odia (low-resource Indian language) speech data $40, odia
- Off-the-shelf speech recognition training datasets (wake words, read sentences, dialogues; Chinese and foreign languages), usage licences sold by Haitian Ruisheng in 2020: 198,951 hours $60, speech recognition corpus
- Off-the-shelf speech synthesis (TTS) training datasets, usage licences sold by Haitian Ruisheng in 2020: 1,029 hours $2,005, tts corpus
- Off-the-shelf speech recognition training datasets (wake words, read sentences, dialogues; Chinese and foreign languages), usage licences sold by Haitian Ruisheng in 2019: 136,411 hours $79, speech recognition corpus
- Off-the-shelf speech synthesis (TTS) training datasets, usage licences sold by Haitian Ruisheng in 2019: 274 hours $3,609, tts corpus
- Off-the-shelf speech recognition training datasets (wake words, read sentences, dialogues; Chinese and foreign languages), usage licences sold by Haitian Ruisheng in 2018: 115,636 hours $76, speech recognition corpus
- Off-the-shelf speech synthesis (TTS) training datasets, usage licences sold by Haitian Ruisheng in 2018: 96 hours $3,040, tts corpus
Asking prices off the shelfper hour of audio
$118
Middle half $65 to $266. 109 figures (+1 open-ended), 1 Jun 2021 to 14 Oct 2026.
The figures
- Nexdata Indonesian spontaneous dialogue smartphone speech, 747 hours from 1,190 native speakers, transcribed with PII redacted (on Datarade) $27, indonesian
- Nexdata Indonesian scripted monologue smartphone speech, 359 hours from 496 speakers, transcribed (on Datarade) $56, indonesian
- REMI Data licensed native-speaker speech recordings in 13+ languages including Yoruba, Amharic and Swahili, WAV 16 kHz with verbatim transcripts and perpetual commercial consent (on Datarade) $120, yoruba
- NaijaVoices suite of speech-text datasets (Hausa, Igbo and Yoruba; 1,800+ hours from 5,000+ speakers): the Supporter tier that adds a waiver for commercial use $2.78, african
- SpeechData.ai Indonesian conversational speech dataset: 2,000 hours of spontaneous two-person conversations recorded in Indonesia, 200 native speakers, dual-channel, time-aligned transcripts $65, indonesian
- SpeechData.ai Vietnamese conversational speech dataset: 2,000 hours of spontaneous two-person conversations recorded in Vietnam, 200 native speakers, dual-channel, time-aligned transcripts $65, vietnamese
- Off-the-shelf speech data, major languages $71
- Recorded speech sold off the shelf $95
- Off-the-shelf call-center audio licensed from existing corpora (AIxBlock guide) $11, call center
- Real customer-service calls in English and Hindi for chatbot and voicebot training (WiserBrand), usage-based licence $25, hindi
- Real US customer-service call recordings, speaker-separated 48 kHz, PII-redacted (WiserBrand, 40,000+ hours available), usage-based licence $25, english
- ALLIES corpus: 900 hours of French broadcast news for diarization evaluation (ELRA catalogue ELRA-S0486) $39, french
- American children's read speech (storybooks, songs), microphone, 50.5 hours from 219 children (ELRA catalogue ELRA-S0468) $639, english
- American English prompted speech recorded on mobile phones, 800 hours from 1,842 speakers (ELRA catalogue ELRA-S0437) $192, english
- American English read speech recorded on mobile phones, 215 hours from 349 speakers (ELRA catalogue ELRA-S0467) $181, english
- Italian Speech Corpus 1 (Appen): 44 hours from 202 speakers, 4 microphone channels, office and public places (ELRA catalogue ELRA-S0147) $382, italian
- Spanish Speech Corpus 1 (Appen): 39 hours from 200 speakers, 4 microphone channels, office and public places (ELRA catalogue ELRA-S0149) $431, spanish
- Brazilian Portuguese prompted speech recorded on mobile phones, 1,044 hours from 2,038 speakers (ELRA catalogue ELRA-S0445) $266, portuguese
- BREF-120: 100 hours of French read speech (ELRA catalogue ELRA-S0067) $168, french
- British children's read speech (textbooks, storybooks), high-fidelity microphone, 55 hours from 201 children (ELRA catalogue ELRA-S0474) $798, english
- British English prompted speech recorded on mobile phones, 831 hours from 1,651 speakers (ELRA catalogue ELRA-S0448) $287, english
- British English read speech recorded on mobile phones, 199 hours from 346 speakers (ELRA catalogue ELRA-S0466) $266, english
- Cantonese Speecon: about 231 hours of adult and child speech over 4 microphone channels in office, car and public places (ELRA catalogue ELRA-S0287) $364, cantonese
- Chinese Kids Speech database (Lower Grade): 87 hours from 184 children aged 6 to 10, smartphone (ELRA catalogue ELRA-S0496) $198, mandarin
- Chinese Kids Speech database (Upper Grade): 72 hours from 161 children aged 10 to 12, smartphone (ELRA catalogue ELRA-S0497) $218, mandarin
- CLE Pakistan Urdu Speech Corpus: 61.62 hours of phonetically rich read Urdu on varied microphones (ELRA catalogue ELRA-S0403) $426, urdu
- Corpus of Spontaneous Japanese (CSJ), 650 hours as listed by ELRA, distributed for NINJAL (ELRA catalogue ELRA-S0488) $74, japanese
- Danish SpeechDat(II) FDB-4000: 242 hours of fixed-network telephone speech from 4,000 speakers (ELRA catalogue ELRA-S0073) $259, danish
- Dutch Polyphone Database: 335 hours of Dutch and Flemish telephone speech (ELRA catalogue ELRA-S0010) $117, dutch
- Egyptian Arabic Speecon: 196 hours of Modern Standard Arabic as spoken in Egypt, adults and children, 4 microphone channels in office, car and public places (ELRA catalogue ELRA-S0308) $429, arabic egyptian
- ESTER 2: about 100 hours of transcribed French broadcast news (ELRA catalogue ELRA-S0338) $280, french
- EthioSpeech: 391 hours of read speech in six Ethiopian languages (Amharic, Tigrigna, Oromo, Somali, Afar, Sidama) (ELRA catalogue ELRA-S0494) $15, ethiopian
- French prompted speech recorded on mobile phones, 769 hours from 1,623 speakers (ELRA catalogue ELRA-S0446) $266, french
- French read speech recorded on mobile phones (France, Canada, Africa), 231.9 hours from 406 speakers (ELRA catalogue ELRA-S0465) $245, french
- English spoken by native German speakers (accented English), mobile phones, 535 hours from 1,162 speakers (ELRA catalogue ELRA-S0442) $319, english
- German prompted speech recorded on mobile phones (generic, in-car, smart-home scripts), 1,796.70 hours from 3,442 speakers (ELRA catalogue ELRA-S0449) $319, german
- German read speech recorded on mobile phones, 211 hours from 327 speakers (ELRA catalogue ELRA-S0473) $245, german
- GlobalPhone Hausa: 8.73 hours of read Hausa speech (ELRA catalogue ELRA-S0347) $462, hausa
- GlobalPhone Tamil: 20 hours of read Tamil speech (ELRA catalogue ELRA-S0205) $34, tamil
- Gram Vaani: 130 hours of Hindi phone recordings from 4,000 mostly rural speakers in Bihar, Jharkhand and Madhya Pradesh (ELRA catalogue ELRA-S0405) $431, hindi
- Greek SpeechDat(II) FDB-5000: 422 hours of fixed-network telephone speech from 5,000 speakers (ELRA catalogue ELRA-S0118) $186, greek
- Hindi prompted speech recorded on mobile phones, 759 hours from 1,425 speakers (ELRA catalogue ELRA-S0452) $170, hindi
- Hindi read speech recorded on mobile phones, 240 hours from 401 speakers, transcribed (ELRA catalogue ELRA-S0463) $160, hindi
- Indonesian read speech recorded on mobile phones, 359 hours from 496 speakers (ELRA catalogue ELRA-S0470) $181, indonesian
- Japanese Kids Speech database (Lower Grade): 116.60 hours from 179 children aged 6 to 9, smartphone (ELRA catalogue ELRA-S0411) $118, japanese
- Japanese read speech recorded on mobile phones, 234.8 hours from 799 speakers (ELRA catalogue ELRA-S0476) $170, japanese
- Korean read speech recorded on mobile phones, 197 hours from 291 speakers (ELRA catalogue ELRA-S0475) $181, korean
- LILA Korean: 81.16 hours of mobile-network telephone speech from 1,000 speakers (ELRA catalogue ELRA-S0295) $716, korean
- MEDIA speech database: 70 hours of French human-machine dialogue (hotel booking scenario) (ELRA catalogue ELRA-S0272) $160, french
- MoveOn Speech and Noise Corpus: 6 hours of British English command speech recorded on motorcycles (ELRA catalogue ELRA-S0370) $1,681, english
- OrienTel Egypt: colloquial Egyptian Arabic over fixed and mobile telephone networks, 50 hours from 750 speakers (ELRA catalogue ELRA-S0221) $1,350, arabic egyptian
- OrienTel Jordan: colloquial Arabic over fixed and mobile telephone networks, 50 hours from 757 speakers (ELRA catalogue ELRA-S0289) $1,350, arabic
- OrienTel Tunisia: colloquial Arabic over fixed and mobile telephone networks, 50 hours from 792 speakers (ELRA catalogue ELRA-S0186) $616, arabic
- Russian prompted speech recorded on mobile phones, 1,002 hours from 1,960 speakers (ELRA catalogue ELRA-S0443) $202, russian
- Spanish emotional speech synthesis database: 4 hours from two professional speakers in neutral and six emotional styles (ELRA catalogue ELRA-S0329) $1,121, spanish
- Spanish read speech recorded on mobile phones (Spain, Mexico, Venezuela), 227 hours, transcribed (ELRA catalogue ELRA-S0464) $202, spanish
- Castilian Spanish SpeechDat(II) FDB-4000: 246.67 hours of fixed-network telephone speech from 4,000 speakers (ELRA catalogue ELRA-S0102) $89, spanish
- Swiss-German SpeechDat(II) FDB-2000: 110 hours of fixed-network telephone speech from 2,000 speakers (ELRA catalogue ELRA-S0105) $387, swiss german
- TC-STAR English training corpus: about 290 hours of European Parliament plenary speeches (ELRA catalogue ELRA-S0251) $35, english
- TC-STAR Spanish baseline female voice: 10 hours of one speaker reading in a noise-reduced room, for speech synthesis (ELRA catalogue ELRA-S0309) $874, spanish
- Thai read speech recorded on mobile phones, 203 hours from 498 speakers (ELRA catalogue ELRA-S0471) $260, thai
- THERADIA WoZ corpus: 86 hours of audiovisual sessions for detecting affect in home cognitive-training (healthcare) (ELRA catalogue ELRA-S0499) $78, french
- TRAD Pashto Broadcast News: 108 hours of transcribed TV and radio news from more than 1,000 speakers (ELRA catalogue ELRA-S0381) $291, pashto
- UK English Speecon: 204 hours from 606 adults and 51 children, 4 microphone channels in office, car and public places (ELRA catalogue ELRA-S0215) $412, english
- Vietnamese prompted speech recorded on mobile phones, 760 hours from 1,751 speakers (ELRA catalogue ELRA-S0432) $181, vietnamese
- Welsh SpeechDat(II) FDB-2000: 170 hours of Welsh telephone speech (ELRA catalogue ELRA-S0075) $119, welsh
- 35,000-hour multilingual conversational speech dataset (Nexdata, 100+ languages) $20,000 / purchase (starts at)
- Arabic podcast audio, dual-channel, 6,024 hours (Infobay.AI on Opendatabay) $4.86, arabic
- English podcast audio, dual-channel, 1,165 hours (Infobay.AI on Opendatabay) $4.86, english
- German call-center recordings, dual-channel, 212 hours (Infobay.AI on Opendatabay) $6.23, german
- Hindi call-center recordings, dual-channel, 1,587,658 hours (Infobay.AI on Opendatabay) $0.52, hindi
- Hindi podcast audio, dual-channel, 11,607 hours (Infobay.AI on Opendatabay) $4.86, hindi
- Mizo (low-resource language of north-east India) call-center recordings, dual-channel, 468 hours (Infobay.AI on Opendatabay) $5.06, mizo
- Nepali call-center recordings, dual-channel, 229,645 hours (Infobay.AI on Opendatabay) $2.12, nepali
- Odia call-center recordings, dual-channel, 12,794 hours (Infobay.AI on Opendatabay) $4.88, odia
- Somali call-center recordings, dual-channel, 1,057 hours (Infobay.AI on Opendatabay) $4.98, somali
- Swahili call-center recordings, dual-channel, 194,331 hours (Infobay.AI on Opendatabay) $1.50, swahili
- Tamil call-center recordings, dual-channel, 15,056 hours (Infobay.AI on Opendatabay) $4.87, tamil
- Saudi Arabic full-duplex customer-service dialogue speech (268 speakers, transcribed), 268 hours (Nexdata on Opendatabay) $442, arabic
- INTERSPEECH 2025 MLC-SLM challenge data: conversational speech in 11 languages, 1,500 hours, curated by Datatang (Nexdata on Opendatabay) $351, multilingual
- Tagalog full-duplex multi-channel spontaneous dialogue speech (smartphone, transcribed), 1,100 hours (Nexdata on Opendatabay) $359, tagalog
- US English real customer-service calls, dual-channel, with verbatim transcripts, 1,000 hours (PissedConsumer on Opendatabay) $25, english
- Hindi real customer-service calls with Hindi-English code-switching, dual-channel, with verbatim transcripts, 1,000 hours (PissedConsumer on Opendatabay) $25, hindi
- Amharic (Ethiopia) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, amharic
- Arabic (Egyptian) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, arabic egyptian
- Arabic (Modern Standard, Saudi Arabia) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $90, arabic
- Bengali (Bangladesh) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, bengali
- English (South African) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $60, english
- English (US) conversational speech dataset, 1,500 hours of two-person conversations from 150 native speakers, transcribed, commercial licence $90, english
- French (DR Congo) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, french
- French (France) conversational speech dataset, 1,000 hours of two-person conversations from 100 native speakers, transcribed, commercial licence $95, french
- German conversational speech dataset, 1,000 hours of two-person conversations from 50 native speakers, transcribed, commercial licence $95, german
- Hausa (Nigeria) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, hausa
- Hindi conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, hindi
- Japanese conversational speech dataset, 500 hours of two-person conversations from 50 native speakers, transcribed, commercial licence $95, japanese
- Korean conversational speech dataset, 500 hours of two-person conversations from 50 native speakers, transcribed, commercial licence $95, korean
- Mandarin Chinese (Taiwan) conversational speech dataset, 1,500 hours of two-person conversations from 150 native speakers, transcribed, commercial licence $70, mandarin
- Portuguese (Brazil) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, portuguese
- Spanish (Mexico) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, spanish
- Spanish (Spain) conversational speech dataset, 1,000 hours of two-person conversations from 100 native speakers, transcribed, commercial licence $85, spanish
- Swahili (Kenya) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, swahili
- Tamil conversational speech dataset, 500 hours of two-person conversations from 50 native speakers, transcribed, commercial licence $75, tamil
- Urdu (Pakistan) conversational speech dataset, 1,000 hours of two-person conversations from 100 native speakers, transcribed, commercial licence $65, urdu
- Yoruba (Nigeria) conversational speech dataset, 2,000 hours of two-person conversations from 200 native speakers, transcribed, commercial licence $65, yoruba
- Defined.ai Data Access Plan, Elite tier: 3,125 hours of off-the-shelf speech data from its marketplace $178, multilingual
- Defined.ai Data Access Plan, Premium tier: 625 hours of off-the-shelf speech data from its marketplace $178, multilingual
- Defined.ai Data Access Plan, Standard tier: 315 hours of off-the-shelf speech data from its marketplace $177, multilingual
- Defined.ai Data Access Plan, Ultimate tier: 1,565 hours of off-the-shelf speech data from its marketplace $178, multilingual
- Corpus of Spontaneous Japanese (CSJ), 661 hours of mostly monologue speech with 7.52 million words of transcripts: a 2-year commercial licence for up to 10 users $4.78, japanese
- Corpus of Spontaneous Japanese (CSJ), 661 hours of mostly monologue speech with 7.52 million words of transcripts: an open-ended commercial licence for up to 10 users $48, japanese
Asking prices off the shelfper record
$1.27 to $1.49too few figures for a median
2 figures, 11 Oct 2026.
The figures
- Sound-effects catalogue for AI training: 350+ libraries, 53,000 files (Krampfstadt Studio), one-off purchase $1.27, sound effects
- Sound-effects catalogue for AI training: 350+ libraries, about 53,000 files (vehicles, machinery, ambiences, foley) from Krampfstadt Studio, on Opendatabay $1.49, sound effects
Asking prices off the shelfper contract year
$22.4Ktoo few figures for a median
1 figure, 11 Oct 2026.
The figures
Asking prices off the shelf (research use only)per hour of audio
$14
Middle half $6.86 to $32. 36 figures, 11 Oct 2026 to 14 Oct 2026.
The figures
- IARPA Babel Vietnamese Language Pack IARPA-babel107b-v0.7 (LDC2017S01): about 201 hours of Vietnamese conversational and scripted telephone speech with transcripts, collected by Appen in 2012 $0.12, vietnamese
- AIDA Ukrainian Broadcast and Telephone Speech: about 156 hours with 1.2 million words of transcripts $13, ukrainian
- AISHELL-1: about 520 hours of Mandarin read speech from 400 speakers on three devices $5.77, mandarin
- AnnoDIFP Session Audio and Transcripts: 438.34 hours of English in-person interview audio with personality scores $6.84, english
- Air Traffic Control Complete: about 70 hours of controller and pilot radio transmissions $16, english
- BOLT CTS CALLFRIEND CALLHOME Egyptian Arabic Audio: about 116 hours of telephone conversations $17, arabic egyptian
- Call My Net 1: 364 hours of conversational telephone speech in Tagalog, Cebuano, Cantonese and Mandarin $6.87, multilingual
- CALLHOME American English Second Edition: about 56 hours of unscripted telephone conversations $54, english
- CALLHOME German Second Edition: about 48 hours of unscripted telephone conversations $63, german
- CALLHOME Japanese Second Edition: about 49 hours of unscripted telephone conversations $61, japanese
- CALLHOME Mandarin Chinese Second Edition: about 38 hours of unscripted telephone conversations $79, mandarin
- CALLHOME Spanish Second Edition: about 38 hours of unscripted telephone conversations $79, spanish
- Fisher English Training Speech Part 1: 984 hours of English conversational telephone speech $7.11, english
- Fisher Levantine Arabic Conversational Telephone Speech: 45 hours (279 calls) $22, arabic levantine
- Fisher Spanish Speech: roughly 163 hours of Spanish telephone speech from 136 speakers $15, spanish
- GALE Phase 2 Arabic Broadcast News Speech Part 1: about 165 hours of Arabic broadcast news $12, arabic
- Global TIMIT Learner Treebank English: about 24 hours of native and non-native (L2) English read speech $31, english
- HKUST Mandarin Telephone Speech Part 1: about 149 hours of Mandarin conversational telephone speech $20, mandarin
- IARPA Babel Zulu Language Pack: about 211 hours of Zulu conversational and scripted telephone speech $0.12, zulu
- ICSI Meeting Speech: about 72 hours of English meeting speech (883 hours of multi-microphone audio) $26, english
- Korean Telephone Conversations Speech: about 44 hours of audio (100 calls) $34, korean
- King Saud University Arabic Speech Database: 590 hours of Arabic speech from 269 speakers $3.39, arabic
- Magic Data Chinese Mandarin Conversational Speech: about 10 hours of conversation from 60 speakers (about 60 hours of multi-device audio) $75, mandarin
- MATERIAL Somali-English Language Pack: about 80 hours of Somali telephone speech with transcripts and translations $3.12, somali
- MATERIAL Swahili-English Language Pack: about 112 hours of Swahili telephone speech with transcripts and translations $2.23, swahili
- Mixer 6 CHiME 8 Transcribed Calls and Interviews: 80 hours of English interviews and telephone speech with transcripts $38, english
- Multi-Language Conversational Telephone Speech 2011, South Asian: about 118 hours of Bengali, Hindi, Punjabi, Tamil and Urdu telephone speech $21, indic
- Multi-Channel WSJ Audio: about 100 hours of British English read speech from 45 speakers, far-field microphone arrays $15, english
- Polish Speech Database: about 280 hours of read Polish speech from 200 speakers $11, polish
- REMIX Telephone Collection: 320 hours of English conversational telephone speech from 358 speakers $7.81, english
- Samromur Children Icelandic Speech 1.0: 131 hours of prompted speech from 3,175 children aged 4 to 17 $1.91, icelandic
- Mandarin-English Code-Switching in South-East Asia (SEAME): about 192 hours from 156 speakers $10, mandarin english
- Switchboard-1 Release 2: about 260 hours of US English conversational telephone speech (543 speakers) $12, english
- TORGO Database of Dysarthric Articulation: about 23 hours of English speech from speakers with cerebral palsy or ALS and controls $52, english
- United Nations Proceedings Speech: about 8,500 hours of proceedings in the six official UN languages $0.59, multilingual
- CSR-I (WSJ0) Complete: about 141 hours of English read speech (123 speakers reading the Wall Street Journal) $11, english
Made to orderper hour of audio
$122
Middle half $90 to $245. 9 figures (+1 open-ended), 1 Jan 2026 to 17 Oct 2026.
The figures
- Custom speech data collection to a buyer's spec (recruited speakers, recorded, transcribed and QA'd), 4 to 12 weeks $122
- Studio-quality text-to-speech (TTS) voice data recorded to order $316, studio tts
- Custom on-site speech and audio recordings made to the buyer's requirements by field teams (Spire Light on Datarade) $250
- Custom paired conversational speech recordings (audio, optionally video) built to project requirements in 80+ languages (Spire Light on Datarade) $90, conversational
- Custom 1,000-hour English call-center speech collection with accent coverage and basic transcription, 12 weeks $126, call center
- Two-speaker spontaneous conversations collected on demand in rare and low-resource languages and accents, with transcripts (Silencio Network) $50, low resource
- Human transcription of English audio, 99%+ accuracy, 12-hour turnaround (Rev) $119, english
- Human transcription of audio and video, 24-hour turnaround (Speechpad) $78, english
- Human-edited transcription, 99%+ accuracy, 1 to 3 day turnaround (TranscribeMe) Starting at $0.79/min, english
- Speech collected to a buyer's spec $245
Cost to produceper hour of audio
$244 to $333too few figures for a median
2 figures, 4 Sep 2025 to 14 Oct 2026.
By segment, where two or more figures exist
- Englishpaid to the people who make it, per hour of work$506 figures
- Germanpaid to the people who make it, per hour of work$442 figures
- Voicepaid to the people who make it, per hour of work$5.00 to $262 figures
- Frenchpaid to the people who make it, per hour of work$18 to $252 figures
- Mandarinpaid to the people who make it, per hour of work$20 to $242 figures
- Bengalipaid to the people who make it, per hour of work$7.50 to $152 figures
- Tts corpussales that happened, per hour of audio$3,0403 figures
- Speech recognition corpussales that happened, per hour of audio$763 figures
- Voice licensingsales that happened, per hour of audio$1.20 to $8.002 figures
- Hausaasking prices off the shelf, per hour of audio$65 to $4622 figures
- Arabicasking prices off the shelf, per hour of audio$4425 figures
- Arabic egyptianasking prices off the shelf, per hour of audio$4293 figures
- Urduasking prices off the shelf, per hour of audio$65 to $4262 figures
- Portugueseasking prices off the shelf, per hour of audio$65 to $2662 figures
- Spanishasking prices off the shelf, per hour of audio$2027 figures
- Mandarinasking prices off the shelf, per hour of audio$1983 figures
- Englishasking prices off the shelf, per hour of audio$19215 figures
- Vietnameseasking prices off the shelf, per hour of audio$65 to $1812 figures
- Koreanasking prices off the shelf, per hour of audio$1813 figures
- Multilingualasking prices off the shelf, per hour of audio$1785 figures
- Germanasking prices off the shelf, per hour of audio$1704 figures
- Frenchasking prices off the shelf, per hour of audio$1609 figures
- Yorubaasking prices off the shelf, per hour of audio$65 to $1202 figures
- Japaneseasking prices off the shelf, per hour of audio$856 figures
- Swahiliasking prices off the shelf, per hour of audio$1.50 to $652 figures
- Indonesianasking prices off the shelf, per hour of audio$604 figures
- Hindiasking prices off the shelf, per hour of audio$458 figures
- Tamilasking prices off the shelf, per hour of audio$343 figures
- Sound effectsasking prices off the shelf, per record$1.27 to $1.492 figures
- Spanishasking prices off the shelf (research use only), per hour of audio$15 to $792 figures
- Mandarinasking prices off the shelf (research use only), per hour of audio$484 figures
- Englishasking prices off the shelf (research use only), per hour of audio$1612 figures
- Arabicasking prices off the shelf (research use only), per hour of audio$3.39 to $122 figures
- Multilingualasking prices off the shelf (research use only), per hour of audio$0.59 to $6.872 figures
- Englishmade to order, per hour of audio$78 to $1192 figures
- Africancost to produce, per hour of audio$244 to $3332 figures
Images
Photo libraries by the image: cents in bulk, dollars with metadata and releases.
Paid to the people who make itper image
$0.54
Middle half $0.29 to $1.31. 8 figures, 5 Feb 2026 to 14 Oct 2026.
The figures
- New RAW photos (or paired images) taken by the contributor for a client's computer-vision training set, Ronia Raw Photo Collection, residents of United States $1.50, united states
- Old photo restoration set: two provided old photos, each restored and paired with written edit instructions, for AI image-editing training data $2.50, image editing
- Project Vertex: a set of 4 progressive edits of a provided real photo (compositing, background swap, object insert or removal, colour change) with written step notes, for AI image-editing training data $1.25, image editing
- Face deduplication collection: neutral selfies, head-pose captures and historical facial photos $0.55, faces
- Face Deduplication collection: a set of 22 to 24 approved photos of oneself (current front-facing and head-pose shots plus older photos from the camera roll), contributors aged 45 or over $0.52, faces
- Photos of everyday objects at home, uploaded with object name and details, for a global AI training data project (gig workers in several countries) $0.28, crowdsourced
- Natural real-world photos for AI training, Zone 1 contributors (India, Philippines, Indonesia, Vietnam, Bangladesh, Pakistan, Nepal), no logos, IDs or plates, full rights held $0.22, south asia
- Smartphone photo collection for an AI training image dataset (tripod, JPEG of 4 MB or more, minimum 250 photos) $0.3, crowdsourced
Paid to the people who make itper hour of work
$5.00 to $35too few figures for a median
2 figures, 11 Nov 2025 to 7 Oct 2026.
The figures
- Annotation of drone and maritime imagery, United States $35, drone
- Polygon segmentation of recycling images $5.00, annotation
Paid to the people who make itper task
$7.50too few figures for a median
1 figure, Jan 2026.
Offered to holdersper image
$0.35
Middle half $0.15 to $2.52. 6 figures, 5 Apr 2024 to 12 Feb 2026.
The figures
- Photos of televisions and displays taken inside the home, existing or newly shot, for a corporate client's image-recognition training set, bought per accepted photo through Imagenavi's Mitsukaru Mall (Japan) $0.13, japan
- Existing camera-roll photos sold through the Kled app (no screenshots, downloads or AI images), as reported by users $0.007, camera roll
- Photos of parked motorbikes and mopeds (fuel-powered two-wheelers), existing or newly shot, for a corporate client's image-recognition training set, bought per accepted photo through Imagenavi's Mitsukaru Mall (Japan) $3.20, japan
- Exterior photos of apartment buildings, public facilities and commercial facilities, existing or newly shot, to train an AI that generates architectural design proposals, bought per accepted photo through Imagenavi's Mitsukaru Mall (Japan) $0.47, japan
- Photographs and artworks licensed through Wirestock for AI training, specific-need requests (top of range) $17, creator
- A photo host's archive of 13 billion photos (Photobucket, in talks) $0.22
Sales that happenedper image
$0.063
Middle half $0.023 to $0.11. 4 figures, 8 May 2023 to 12 Jun 2025.
The figures
- Shutterstock's 2023 AI licensing revenue spread across its 771 million image archive $0.14, stock library
- Photos bought by Adobe from contributors to fill gaps in its AI training set $0.098, creator
- Most of Freepik's roughly 200 million image archive licensed to two large tech companies $0.028
- Shutterstock Contributor Fund payout (AI data-licensing share), median per image in the contributor's portfolio, May 2023 payout, survey of 58 contributors $0.0069, stock library
Sales that happenedper contract
$378K to $35.4Mtoo few figures for a median
2 figures (+3 open-ended), 6 Apr 2024 to 28 Jul 2026.
The figures
- De-identified medical imaging records and clinical reports for a foundation model (OneMedNet, initial agreement) in excess of $11.5 million, medical
- Aerometrex 3D city datasets and LiDAR licensed off the shelf to two AI firms (Zeromatter, Neara) for training $378K, 3d lidar
- Two multiyear licences of Getty Images content (a major social platform and a large AI company) approximately $65,000,000 total deal value; about $40,000,000 recognized in Q4 2025
- Shutterstock Data, Distribution, and Services revenue (library and metadata licensing, including to AI developers), fourth quarter of 2025, all buyers $30.7 million, down $7.1 million or 19% on Q4 2024
- Shutterstock image-library licences to large tech buyers $35.4M
Sales that happenedper contract year
$30.4Mtoo few figures for a median
1 figure, Nov 2023.
Asking prices off the shelfper image
$1.71
Middle half $0.27 to $5.27. 24 figures (+9 open-ended), 5 Apr 2024 to 11 Oct 2026.
The figures
- Unidata Medical Masks (face-mask detection) Dataset, 167,880 face images of 41,970 people in 4 mask states Starts at $10,000 / purchase, faces
- Unidata Kids & Teens Aging Dataset (ages 7 to 15), 10,214 facial images of 1,022 minors Starts at $30,000 / purchase, faces
- Unidata Car License Plate Detection Dataset, 3,809,704 annotated road-camera images across 93 countries Starts at $30,000 / purchase, driving
- Unidata Open Palm Hand Images Dataset, 500,000 annotated images from 50,000 people Starts at $10,000 / purchase, hands
- Unidata Selfie with ID Dataset, 69,435 selfie and ID-photo images of 4,629 people from 85 countries Starts at $10,000 / purchase, faces
- Unidata Synthetic Passports Dataset, 100,000 AI-generated passport images with MRZ metadata Starts at $30,000 / purchase, document images
- 20K+ Infectious Dermatology Images with Prescriptions Dataset (JPG images paired with PDF prescriptions) $9.81, medical
- 187K+ X-Ray Images Dataset (DICOM, 79K+ anonymized patients), Infobay.Ai on Opendatabay $4.88, medical
- 193K+ X-Ray Images Dataset (JPG, 119K+ anonymized patients), Infobay.Ai on Opendatabay $5.05, medical
- Facial Skin Condition Dataset, 639 images of 213 people (acne, blackheads, redness, under-eye bags) Starts at $1,000 / purchase, medical
- Hair Loss Dataset, 6,020 labeled scalp images of 1,160 people (alopecia classification and segmentation) Starts at $2,000 / purchase, medical
- H&E pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 1–99 slide tier $5.00, medical
- H&E pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 100–999 slide tier $4.00, medical
- H&E pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 1000–9999 slide tier $3.00, medical
- H&E pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 10000+ slide tier $2.00, medical
- IHC and other pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 1–99 slide tier $40, medical
- IHC and other pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 100–999 slide tier $32, medical
- IHC and other pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 1000–9999 slide tier $27, medical
- IHC and other pathology whole slide images from the HistAI DataHub, perpetual commercial licence, 10000+ slide tier $23, medical
- Single archive high-resolution optical satellite image, SkyFi marketplace Starting at $ 15, satellite
- AKYS Pathfinder A2 Motorway 360-degree imagery, Lithuania, 2,081 11K panoramas $0.22, street imagery
- AKYS Pathfinder Vilnius City 360-degree street imagery, 16,318 11K equirectangular panoramas with per-image pose, faces and plates blurred $0.044, street imagery
- MOYU Solutions Drone Park Vehicle Congestion Dataset, 20,000 1080p drone images, 30 GB $0.53, drone
- Matthew Britton Illustration, full Lineartestpilot stock illustration portfolio, 850,000 hand-drawn images x 3 formats (2.5M files) $0.093, illustration
- Matthew Britton Illustration, 10,000 hand-drawn vector images x 3 formats (JPG, SVG, PNG) subset of the Lineartestpilot portfolio $0.4, illustration
- Infobay.AI 1.2M+ Infographic Image Dataset (price for initial batch of 500,000 infographic images) $0.97, document images
- Infobay.AI 873K+ Scalable Vector Graphics (SVG) Image Dataset (icons, illustrations, logos, diagrams, UI elements) $0.29, illustration
- Alessandro Grandini Jewelry & Fashion Accessories Product Dataset, 5,000 studio JPGs on white, 3000 to 4000 px $0.026, product photos
- Alessandro Grandini Facial Expressions Close-Up Dataset, 2,000 high-resolution (3000 to 6000 px) face close-up photos $0.032, faces
- JoinAI Drone-Based Rooftop Leaf Detection Image Dataset, 20,000+ annotated drone photos, 4K and 1080p $0.53, drone
- Bulk photo datasets licensed for AI training (DatasetShop published rates) $0.05
- Images of nudity, which need special handling (Defined.ai) $5.92, sensitive
- Images sold through a data marketplace (Defined.ai's quoted rate) $1.41
Asking prices off the shelfper record
$1.13
Middle half $1.00 to $1.88. 3 figures (+1 open-ended), 7 Aug 2026 to 11 Oct 2026.
The figures
- Unidata Anti-Spoofing Real Videos Dataset, 87,340 files (video clips and selfies) of 43,700 people, consented crowdsourcing Starts at $15,000 / purchase, faces
- Medical Imaging & Radiology Dataset (CT, X-ray, MRI and other modalities with diagnostic reports, DICOM) $2.63, medical
- Nexdata 1.51 Million Sets of single-image and multi-image fusion image editing data (before/after pairs with instructions) $0.87, image editing
- Nexdata 116,048 sets of 3D hand pose gesture data (RGB hand mask, 16-bit depth, keypoints, mesh) $1.13, hands
Asking prices off the shelfper dataset
freetoo few figures for a median
1 figure (+1 open-ended), Oct 2026.
The figures
- Multi-View Drone Imagery & 3D Scenes, multi-country computer vision dataset Starts at $10,000 / purchase, drone
- The Cancer Imaging Archive (NCI) collections (CT, MRI, PET, pathology) under CC BY 3.0 or 4.0 free, medical
Made to orderper image
$0.44
Middle half $0.064 to $0.91. 4 figures (+1 open-ended), 29 Mar 2026 to 11 Oct 2026.
The figures
- Custom image classification dataset, focused single-domain (5,000 to 20,000 images) $0.8, classification
- Single newly tasked high-resolution optical satellite image, SkyFi Starting at $ 200, satellite
- Bounding box image annotation $0.077, annotation
- Image classification labeling $0.022, annotation
- Semantic (pixel-level) segmentation labeling $1.22, annotation
Made to orderper task
$100
Middle half $78 to $200. 3 figures, Jul 2017.
The figures
- Radiologist read of a research body CT exam (abdomen and pelvis, non-contrast), Penn CBI FY18 rate schedule $100, medical
- Neuroradiologist read of a research CT exam (non-contrast), Penn CBI FY18 rate schedule $55, medical
- Structural MRI segmentation and analysis of one research MRI dataset, Penn CBI Neuroimaging Core $300, medical
By segment, where two or more figures exist
- Image editingpaid to the people who make it, per image$1.25 to $2.502 figures
- Facespaid to the people who make it, per image$0.52 to $0.552 figures
- Crowdsourcedpaid to the people who make it, per image$0.28 to $0.32 figures
- Japanoffered to holders, per image$0.473 figures
- Stock librarysales that happened, per image$0.0069 to $0.142 figures
- Medicalasking prices off the shelf, per image$5.0511 figures
- Droneasking prices off the shelf, per image$0.532 figures
- Illustrationasking prices off the shelf, per image$0.293 figures
- Street imageryasking prices off the shelf, per image$0.044 to $0.222 figures
- Annotationmade to order, per image$0.0773 figures
- Medicalmade to order, per task$1003 figures
Text and archives
Books, reporting and reference text, licensed by the title or the word.
Paid to the people who make itper hour of work
$52
Middle half $40 to $70. 3 figures, 21 Aug 2026 to 15 Oct 2026.
The figures
- Humanities expert work on Mercor (standing talent network; band its listings have posted at) $52, writing
- AI language expert, French (native speakers with teaching experience) on Alignerr $27, french
- News analysts, reporters and journalists evaluating AI-generated news content (4+ years' experience) $88, journalism
Offered to holdersper book or title
$355
Middle half $280 to $1,428. 3 figures (+6 open-ended), 12 Dec 2024 to 18 Nov 2025.
The figures
- High end of the per-book AI licensing offers Created by Humans says it has seen for authors' books (founder's statement, no buyer named) tens of thousands per book
- One-time AI training royalty Arcadia Publishing offered a Belt Publishing author (Rethinking Fandom, 2022) in its second round of sublicensing to an unnamed major technology company, opt-out by default $355, trade book
- Royalty Bloomsbury offers academic and professional authors who opt in to AI training licences of their backlist titles a royalty rate of 20 percent, scholarly book
- Author's expected return when Johns Hopkins University Press licenses a backlist title to an unnamed LLM developer (opt-out programme, about 3,000 backlist titles) less than $100 per title per license, scholarly book
- One-time AI training royalty Arcadia Publishing offered the author of a local-history title (Images of Baseball: Kane County Cougars) for a sublicence to an unnamed major technology company, first round, opt-out by default $205, trade book
- Black Inc (Melbourne trade publisher): contract variation asking its authors to grant AI training rights to their books, author's share of the licence receipts net receipts split with the author 50/50, book
- Author's half of HarperCollins' per-title AI training fee (select nonfiction backlist, three-year licence, buyer reported as Microsoft) $2,500, book
- Royalty Cambridge University Press pays authors who opt in when their work is used in a generative AI licensing deal a royalty of 20%, scholarly book
- Share of AI licensing income that some academic presses offer their authors per title, as reported by the Authors Guild (presses not named) a mere 25 percent, scholarly book
Offered to holdersper contract
Open-ended only
0 figures (+1 open-ended), Dec 2023.
The figures
- Apple's reported multiyear offers to news publishers (Conde Nast, NBC News, IAC) for archives to train its generative AI Reported offers of 50M-plus per publisher, news
Sales that happenedper book or title
$2,750
Middle half $881 to $3,082. 10 figures (+2 open-ended), 13 Sep 2022 to 24 Jul 2026.
The figures
- Per-title damages the Seoul Central District Court (appeal) set for e-books copied without permission into the National Institute of Korean Language written corpus, 78 rights holders of Communication Books titles against Woongjin and Woongjin Booxen, 2026 전자책 정가의 70%에 인쇄 부수 5부 (70% of the e-book list price, times 5 copies, times the plaintiff's share), korean
- Bartz v. Anthropic class settlement: the $1.5 billion fund divided over the 482,460 works on the final Works List (books downloaded from pirate libraries) $3,109, settlement
- Bartz v. Anthropic class settlement: the fund left after the court-awarded fees, expenses, cost reserve and service awards, divided over the 440,490 works claimed by April 2026 $3,127, settlement
- Books in the Bartz v. Anthropic class settlement (pirated copies used for training) $3,000, settlement
- What one academic author received as his share when a small, ethical academic publisher licensed one of his monographs to an unnamed AI training company (opt-in by tick box, publisher keeps 50% of the total fee, buyer undisclosed under NDA) $674, scholarly book
- Author's default half of the per-work award in the Bartz v. Anthropic settlement (trade and university press books on the Works List) $1,500, book
- AI training royalty per title paid to History Press (Arcadia Publishing) authors who stayed in Arcadia's 2025 sublicence to an unnamed technology company; one author took it for three titles $340, trade book
- Bartz v. Anthropic class settlement: the payment Anthropic owes for each class work added to the Works List above 500,000 $3,000, settlement
- What one Taylor & Francis author says they earned for their book under the publisher's AI training licence to Microsoft $97, scholarly book
- HarperCollins AI training licence fee actually paid to one author for one nonfiction backlist title (Writing Without Bullshit, 2016): three-year opt-in licence, author signed and received a cheque about a month later $2,500, book
- One nonfiction book title, licensed for training for three years (HarperCollins) $5,000
- Per-title settlement Woongjin Booxen set with the publishers whose Bookopia e-books it used without permission in the National Institute of Korean Language written corpus (an AI and language-research corpus), Korea, 2022 전자책 정가의 70%의 3copy (70% of the e-book list price, times 3 copies), korean
Sales that happenedper contract year
$16M
Middle half $7.1M to $50M. 9 figures (+2 open-ended), Feb 2024 to 23 Sep 2026.
The figures
- Times of India and Economic Times content (Bennett, Coleman & Co) surfaced in ChatGPT under an OpenAI partnership $5M, news
- Indian Express Group content licensed to OpenAI for ChatGPT $3M, news
- Wiley scholarly and learning content licensed for AI (all licensees, fiscal year to 30 April 2026) $49.1 million (prior year $40 million)
- News Corp US and UK news content, including archives, licensed to Meta for AI up to US$50 million a year, at least three years, news
- New York Times news, cooking and Athletic content licensed to Amazon for AI training and summaries $22.4M, news
- Dotdash Meredith (now People Inc.) magazine and web content licensed to OpenAI (2024 deal) $16M, magazines
- News Corp's current and archived journalism licensed to OpenAI for five years (2024 deal) $50M, news
- Reddit's user posts licensed to OpenAI for training and display (2024 deal) $70M, forum posts
- Taylor & Francis academic books and journals licensed to Microsoft for AI (first year of the deal) $10M, academic
- Financial Times journalism licensed to OpenAI (2024 deal) $7.1M, news
- Reddit's user posts licensed to Google for AI training and access (2024 deal) $60M, forum posts
Sales that happenedper contract
$23M to $203Mtoo few figures for a median
2 figures (+5 open-ended), Dec 2023 to 30 Jul 2026.
The figures
- Reddit other revenue (mostly data licensing, plus products sold directly to users), second quarter of 2026, all licensees $43 million, up 24% year over year
- Total damages awarded on appeal to 78 rights holders of Communication Books titles copied into the National Institute of Korean Language written corpus (Woongjin and Woongjin Booxen), 2026 약 344만 원 (about 3.44 million won), korean
- Publisher content used in Perplexity's Comet browser and assistant (revenue share) 80% of $5 monthly Comet Plus revenue; $42.5 million pool for early partners, news
- Taylor & Francis (Informa) incremental non-recurring AI data licensing revenue in the first half of 2025, including a new 2025 contract with a different, undisclosed LLM customer (academic books and content) about a £15 million incremental, non-recurring data licensing revenue number in the first half, scholarly book
- Wiley's learning and research content licensed to tech companies for AI, total earned to date by August 2024 $23M, academic
- Reddit's data licences signed in January 2024 (access to search and analyse Reddit data), aggregate of several contracts $203M, forum posts
- Axel Springer news content licensed to OpenAI (December 2023 deal) a reported figure in the tens of millions of euros, news
Asking prices off the shelfper word
$0.001
Middle half $0.00011 to $0.0025. 9 figures (+4 open-ended), 5 Apr 2024 to 11 Oct 2026.
The figures
- 60M+ words of editorially curated magazine text (about 90 issues, English, Hindi and Indian regional languages), 100% cleared IP $150,000, indic
- Amharic-English bilingual news and legal corpus (232,653 Amharic and 291,701 English words) $0.0085, amharic
- CRATER corpus (aligned telecom manuals, 1,000,000 tokens per language in English, French and Spanish) $0.000037, parallel
- Dutch PAROLE Distributable Corpus (3 million words of books, newspapers and periodicals) $0.00093, dutch
- La Depeche de Kabylie corpus (about 1,570,000 words of Amazigh newspaper text) $0.00011, amazigh
- Karl May Korpus (the works of the German author Karl May, about 1.6 million words) $0.0025, book
- MULTEXT JOC corpus (written questions and answers of the EU Official Journal, about 5 million words in five languages) $0.0011, parallel
- Wojood nested Arabic named-entity corpus (about 550,000 tokens, MSA and dialect, annotated) $0.0076, annotated
- 1,000+ Bengali non-STEM textbooks, 74M+ words (PDF), commercial AI licence $108,600, bengali
- 140+ Sanskrit non-STEM textbooks, 8M+ words (PDF), commercial AI licence $11,400, sanskrit
- 1,000+ Urdu non-STEM textbooks, 57M+ words (PDF), commercial AI licence $81,700, urdu
- English Gigaword Fifth Edition, commercial reduced-licence fee (4.03 billion words of newswire) $0.00000074, news
- General text sold through a data marketplace (Defined.ai's quoted rate) $0.001
Asking prices off the shelfper record
$0.013
Middle half $0.0048 to $0.018. 6 figures, 11 Oct 2026.
The figures
- Quora questions and answers dataset (10.2M+ records, 21 fields) scraped and sold by Bright Data $0.0025, forum
- Wikipedia articles dataset (9.2M+ records, 11 fields) scraped and sold by Bright Data $0.0025, reference
- Annotated tweet corpus in Arabizi, French and English (585,163 tweets, opinion and theme annotations) $0.019, social
- Chinese-Vietnamese Parallel Corpus (200,000 sentence pairs, average 15 words per sentence) $0.012, parallel
- English-Vietnamese Parallel Corpus (1,000,000 sentence pairs, average 20 words per sentence) $0.013, parallel
- 5 million real consumer reviews and complaints with intent labels (PissedConsumer data) $0.15, user generated
Asking prices off the shelfper article
$0.22
Middle half $0.21 to $0.22. 5 figures, 11 Oct 2026.
The figures
- Al-Hayat Arabic newspaper corpus (42,591 articles in 7 domains, 18.6M tokens) $0.038, arabic
- Archives of El Mundo newspaper, 2020 to 2022 (45,658 Spanish articles, El Mundo and sister titles) $0.22, news
- Archives of El Mundo newspaper, year 2020 (15,073 Spanish articles) $0.22, news
- Archives of El Mundo newspaper, year 2021 (14,461 Spanish articles) $0.23, news
- Archives of El Mundo newspaper, year 2022 (16,124 Spanish articles) $0.21, news
Asking prices off the shelfper book or title
$78
Middle half $60 to $80. 4 figures (+6 open-ended), 11 Oct 2026.
The figures
- 100,000+ rights-cleared academic and professional e-books (PDF, EPUB; English, Hindi and other languages) from Arts & Science Academic Publications $53,514,000, book
- 18,000+ rights-cleared STEM and non-STEM EPUB e-books from Indian publishers, licensed by Arts & Science Academic Publications $9,473,700, book
- MEMODATA tagged French text: a tagged corpus of 50 books (19th-century novels by Balzac, Hugo, Stendhal and others, plus legal texts) $86, book
- 76,087 full-text books in 10 Indian languages (29.3 million pages, about 11.7 billion tokens) $6.92, indic
- 6,400+ Arabic non-STEM textbooks, 336M+ words (PDF), commercial AI licence $506,800, arabic
- 6,000+ English non-STEM textbooks, 521M+ words (PDF), commercial AI licence $493,400, textbook
- 8,939 English STEM textbooks, 855M+ words (PDF), commercial AI licence $78, textbook
- 2,800+ Hindi non-STEM textbooks, 250M+ words (PDF), commercial AI licence $222,100, hindi
- 5,169 Indonesian STEM textbooks, 208.30M+ words (PDF), commercial AI licence $78, indonesian
- 300+ legal textbooks, 32M+ words across 80K+ pages (PDF), commercial AI licence $26,600, legal
Asking prices off the shelfper dataset
$562 to $2,125too few figures for a median
2 figures, 11 Oct 2026 to 17 Oct 2026.
The figures
- Silicon-DPO Platinum, The Reasoning-Code Dataset: 5,001 preference pairs with think-trace and self-corrected code, verified by Python unit tests in a sandbox (coding 35%, finance 20%, STEM 15%; JSONL), Commercial / Startup licence, sold on Gumroad $562, preference pairs
- LDC for-profit standard membership: access to 16 corpora published in the membership year (text and speech) $2,125, for profit
Asking prices off the shelfper document or page
$4.88 to $49too few figures for a median
2 figures, 11 Oct 2026.
Asking prices off the shelfper token
$0.00000082 to $0.0000011too few figures for a median
2 figures, 11 Oct 2026.
The figures
- High-quality Arabic web corpus (FineWeb-edu Common Crawl text machine-translated by an LLM), 13.3M documents, 14.9B tokens $0.0000011, common crawl
- High-quality Spanish, Arabic and Norwegian web corpus (FineWeb-edu Common Crawl text machine-translated by an LLM), about 40M documents, 45B tokens $0.00000082, common crawl
Asking prices off the shelf (research use only)per word
$0.00013
Middle half $0.000003 to $0.00067. 21 figures (+4 open-ended), 11 Oct 2026.
The figures
- An-Nahar Arabic newspaper text, 1995 to 2000 (6 years, 45,000 articles and 24 million words a year) $0.000037, arabic
- ECI/MCI European Corpus Initiative Multilingual Corpus I (over 98 million words in most major European languages) 50.00 €, multilingual
- Monolingual Greek corpus (1 million words of 1996 Eleftherotipia newspaper articles) $0.00067, greek
- ICE-GB, British component of the International Corpus of English (a million words of spoken and written English, fully parsed) $0.0017, annotated
- PAROLE French corpus (20.1 million words: Le Monde, CNRS Editions books, periodicals) $0.00024, french
- PRESS 65 (one million running words of 1965 Swedish newspapers, annotated to sentence level) $0.022, swedish
- TIDES Extraction (ACE) 2003 Multilingual Training Data (about 231,000 words in Arabic, Chinese and English annotated for entities and relations) $0.013, annotated
- Annotated English Gigaword (nearly ten million newswire documents, over four billion words, with parses, named entities and coreference) $6,000.00 Non-Member, news
- AQUAINT Corpus of English News Text (roughly 375 million words: AP, NYT and Xinhua newswire) $0.0000008, news
- Arabic Gigaword first edition (about 1 million news documents, 400 million words of Arabic) $0.0000075, arabic
- Chinese English News Magazine Parallel Text (20 million Chinese characters with 9 million words of English translation) $0.00028, parallel
- Chinese Gigaword first edition (about 1 billion words of Mandarin news text) $0.000003, chinese
- Chinese Gigaword Second Edition (about 1.3 billion words of Chinese newswire) $0.0000027, chinese
- Chinese News Translation Text Part 1 (474,000 Chinese characters with commissioned English translations, about 285,000 words) $0.007, parallel
- Chinese Treebank 9.0 (about two million words of annotated and parsed Chinese text) $0.00015, annotated
- English Gigaword first edition (about 1.8 billion words of English newswire, 4.1 million documents) $0.0000017, news
- English Gigaword Second Edition (about 2.3 billion words of English newswire) $0.0000013, news
- English Gigaword Fifth Edition (4.03 billion words, 9.88 million newswire documents, seven English news services, 1994 to 2010) $0.0000015, news
- English Web Treebank (254,830 words of weblogs, newsgroups, email, reviews and Q&A, hand-annotated for syntax) $0.00069, annotated
- French Gigaword Third Edition (863 million words, 3.16 million newswire documents) $0.0000052, french
- ISI Arabic-English Automatically Extracted Parallel Text (1,124,609 sentence pairs, about 31M English words) $0.00013, parallel
- ISI Chinese-English Automatically Extracted Parallel Text (558,567 sentence pairs, about 16M English words) $0.00025, parallel
- Prague Czech-English Dependency Treebank 2.0 (over two million words in close to 100,000 sentences) $100.00 Non-Member, parallel
- Spanish Gigaword Second Edition (1.09 billion words, 3.7 million newswire documents) $0.0000037, spanish
- Treebank-2 (over 1.6 million words of hand-parsed Dow Jones and other text) $1,500.00 Non-Member, annotated
Asking prices off the shelf (research use only)per dataset
$55too few figures for a median
1 figure, 17 Oct 2026.
Asking prices off the shelf (research use only)per article
$0.00055too few figures for a median
1 figure, 11 Oct 2026.
Made to orderper record
$0.02too few figures for a median
1 figure, 14 Oct 2026.
The figures
Made to orderper book or title
$129too few figures for a median
1 figure, May 2019.
Cost to produceper token
$0.0000024too few figures for a median
1 figure, 11 Oct 2026.
The figures
By segment, where two or more figures exist
- Trade bookoffered to holders, per book or title$205 to $3552 figures
- Settlementsales that happened, per book or title$3,0554 figures
- Booksales that happened, per book or title$1,500 to $2,5002 figures
- Scholarly booksales that happened, per book or title$97 to $6742 figures
- Forum postssales that happened, per contract year$60M to $70M2 figures
- Newssales that happened, per contract year$7.1M to $50M2 figures
- Newssales that happened, per contract year$5M3 figures
- Parallelasking prices off the shelf, per word$0.000037 to $0.00112 figures
- Parallelasking prices off the shelf, per record$0.012 to $0.0132 figures
- Newsasking prices off the shelf, per article$0.224 figures
- Common crawlasking prices off the shelf, per token$0.00000082 to $0.00000112 figures
- Annotatedasking prices off the shelf (research use only), per word$0.00124 figures
- Parallelasking prices off the shelf (research use only), per word$0.000264 figures
- Frenchasking prices off the shelf (research use only), per word$0.0000052 to $0.000242 figures
- Arabicasking prices off the shelf (research use only), per word$0.0000075 to $0.0000372 figures
- Chineseasking prices off the shelf (research use only), per word$0.0000027 to $0.0000032 figures
- Newsasking prices off the shelf (research use only), per word$0.00000144 figures
Datasets, tasks and environments
What labs train agents in: a simulated workplace, tasks inside it and a grader for each.
Paid to the people who make itper environment
$1,500
Middle half $1,200 to $2,500. 29 figures, Oct 2025 to 27 Oct 2025.
The figures
- Open-source RL environment built for Prime Intellect's Environments Hub (bounty) $707
- AndroidWorld mobile-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty) $1,000, computer use
- AppWorld (Ai2) reimplemented as an environment: an agent operating a world of simulated everyday apps through APIs (Prime Intellect Environments Hub, application-only bounty) $1,000, app simulator
- BFCL-v3 (Berkeley Function Calling Leaderboard) reimplemented as an environment: tool and function calling (Prime Intellect Environments Hub, application-only bounty) $500, tool use
- Biomni-R0 (Stanford) reimplemented as an environment: a biomedical research agent (Prime Intellect Environments Hub, application-only bounty) $1,500, biomedical
- MCP backend simulator and agent environment for Box (Prime Intellect Environments Hub, application-only bounty) $2,500, app simulator
- Environment synthesised and validated from a textbook: CLRS, Introduction to Algorithms (Prime Intellect Environments Hub, application-only bounty) $1,000, textbook tasks
- ComputerRL (Zhipu) computer-use RL setup reimplemented as an environment (Prime Intellect Environments Hub, application-only bounty) $5,000, computer use
- Software-library environment for agent tasks in Django (Prime Intellect Environments Hub, application-only bounty) $3,000, coding
- MCP backend simulator and agent environment for Figma (Prime Intellect Environments Hub, application-only bounty) $2,000, app simulator
- MCP backend simulator and agent environment for Intercom (Prime Intellect Environments Hub, application-only bounty) $1,200, app simulator
- Mind2Web web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty) $2,000, computer use
- MLE-Bench (OpenAI) reimplemented as an environment: ML engineering competitions for an agent (Prime Intellect Environments Hub, application-only bounty) $1,200, ml engineering
- MCP backend simulator and agent environment for Monday.com (Prime Intellect Environments Hub, application-only bounty) $1,500, app simulator
- nanoGPT speedrunning environment: an agent trains a small language model as fast as it can (Prime Intellect Environments Hub, application-only bounty) $2,000, ml engineering
- Vercel's Next.js eval suite implemented as an environment (Prime Intellect Environments Hub, application-only bounty) $3,000, coding
- MCP backend simulator and agent environment for Notion (Prime Intellect Environments Hub, application-only bounty) $1,500, app simulator
- OpenHands agent framework integration with the Environments Hub (Prime Intellect Environments Hub, application-only bounty) $2,000, framework integration
- OSWorld computer-use benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty) $1,500, computer use
- PaperBench (OpenAI) reimplemented as an environment: an agent replicates research papers (Prime Intellect Environments Hub, application-only bounty) $1,500, research replication
- Software-library environment for agent tasks in PyTorch (Prime Intellect Environments Hub, application-only bounty) $5,000, coding
- Shop-R1 reimplemented as an environment: an agent simulating online shopping behaviour (Prime Intellect Environments Hub, application-only bounty) $2,500, web agent
- MCP backend simulator and agent environment for Stripe (Prime Intellect Environments Hub, application-only bounty) $1,500, app simulator
- SWE-bench Verified coding benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty) $3,500, coding
- TAU2 (Sierra) extended training set generation pipeline: customer-service tool-use tasks generated for training (Prime Intellect Environments Hub, application-only bounty) $1,500, tool use
- WebArena web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty) $2,000, computer use
- MCP backend simulator and agent environment for Webflow (Prime Intellect Environments Hub, application-only bounty) $2,500, app simulator
- WebVoyager web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty) $1,200, computer use
- MCP backend simulator and agent environment for Zapier (Prime Intellect Environments Hub, application-only bounty) $2,000, app simulator
Paid to the people who make itper hour of work
$89
Middle half $64 to $101. 12 figures, 20 Nov 2023 to 15 Oct 2026.
The figures
- Business operations expert work on Mercor, including briefs and reference answers for simulated work environments (standing talent network band) $104
- Machine learning engineer building scenarios and worlds in an RL environment platform (Terac contract) $90
- Experienced QuickBooks user building RL environments and writing their tasks for accounting and sales workflows (Terac) $110
- Evaluator running everyday errands with two AI agents, approving key steps, screen-recording and scoring (micro1, US) $34
- Reproducible physical-science task an AI agent must complete in a Linux terminal, Turing Physical Sciences Expert $43, terminal task
- Auditor of repository-level software-engineering benchmark tasks (reference patches, test harnesses) for Mercor $79
- Engineer building the pipeline that captures and scrubs coding-agent (Copilot CLI) traces for a lab dataset (Turing contract) $100
- Human test-taker calibrating ARC-AGI-2 tasks in a 90-minute session $88
- Software engineering intern at Mechanize designing RL tasks (prompt, environment and automated grader) $150
- Human baseliner timing software and research tasks for METR's HCAST benchmark $71
- Crowd annotator demonstrating web tasks for the Mind2Web agent dataset $10
- PhD-level expert writing and validating GPQA benchmark questions $95
Paid to the people who make itper task
$300
Middle half $175 to $650. 11 figures (+2 open-ended), 17 Feb 2025 to 16 Oct 2026.
The figures
- Authoring one hard terminal-based task with an automated verifier for Terminal-Bench (coding, machine learning, security, systems, hardware, operations and sciences), Outlier with Scale AI, terminal-bench expert page Up to $1,800 USD/task, terminal bench task
- Terminal-style coding task that stumps frontier models (Outlier, run with Scale AI) Up to $1,350 USD/task
- Corporate finance professional building RL environments from financial models and narrated walkthroughs (Terac) $1,000
- Agentic genomics task with reference solution for GeneBench Pro, Turing Bioinformatics / Computational Genomics Expert (October posting) $150, computational biology
- Scientific-computing task (main problem, subproblems, reference solution, tests) authored from the researcher's own work, AfterQuery Scientific Researcher (Computational Task Author) $200, scientific computing
- Terminal-based scientific task for Terminal Bench Science (environment, expert solution, automated tests), Turing Scientific Computing Expert $300, terminal task
- Python backend replica of a SaaS tool (Slack, Linear, Jira, Notion, Gmail) or a long-horizon agent task with rubric, Turing frontier AI data initiative $300, saas environment
- Agentic genomics task with reference solution for GeneBench Pro, Turing Bioinformatics / Computational Genomics Expert (September posting) $260, computational biology
- Research-grade scientific-computing evaluation task authored from the researcher's own computational work, AfterQuery Scientific Computing Research Expert $300, scientific computing
- One computer-use agent task built end to end (multi-step Linux desktop scenario, files, instruction, Python setup and scoring script), Careerflow AI/ML Software Engineer, Task Creator (RL Environments) $50, computer use environment
- Reviewing Android agent action sequences and screenshots, pass or fail, on the contributor's own phone (micro1) $0.22
- Human expert baseline attempt on an 8-hour AI R&D task (METR RE-Bench) $1,855
- Real freelance software tasks from Upwork used as the SWE-Lancer eval set $1,265, coding
Paid to the people who make itper contract year
$315K
Middle half $281K to $385K. 4 figures, 21 Sep 2025 to 11 Oct 2026.
The figures
- Software engineer at Mechanize, hired to build environments and evals for frontier coding agents $346K
- Member of Technical Staff, Deeptune Environments (Mercor), building RL training environments for labs $283K
- Software Engineer (RL Environments) at AfterQuery $275K
- Software engineer salary at Mechanize for building RL environments (2025 report) $500K
Paid to the people who make itper demonstration
$200too few figures for a median
1 figure, 14 Oct 2026.
Paid to the people who make itper question
$1,581too few figures for a median
1 figure (+1 open-ended), 15 Sep 2024 to 11 Jun 2025.
The figures
Sales that happenedper contract
Open-ended only
0 figures (+1 open-ended), 12 Jan 2026.
The figures
- Exclusive rights instead of shared rights Roughly 4 to 5 times the non-exclusive price
Made to orderper contract
$722K
Middle half $387K to $1.1M. 9 figures (+1 open-ended), 15 Feb 2024 to 16 Oct 2026.
The figures
- Adding a robot model not yet supported (URDF or MJCF and, where needed, dynamics identified from hardware) to a custom robotics RL environment $3,162, robotics sim
- RL environment supply contract between a vendor and a smaller (neo) lab $387K
- Agentic evaluation interface (harness, scaffolding, orchestration) for evaluating general-purpose AI models on multi-step tasks in digital environments such as a browser, command line or full operating system, built to order for the EU AI Office over 24 months (Lot 6 of 'Technical assistance for AI safety') $1.1M, agent eval harness
- Administration of the UK AI Safety Institute's fund for sourcing external evaluations (programs that test frontier models' dangerous capabilities), awarded to Smarter Society $3M, eval tasks
- Private capture-the-flag evaluations designed and built to order to test an AI model's offensive cyber capabilities (UK AI Security Institute, prj_3902 WP 2) $722K, cyber ctf
- Anthropic's discussed spend on RL environments bought from vendors over the following year more than $1 billion
- Cyber range environment for an enterprise network built to order for AI agent evaluation: tasks inside it, a solver per task, an automated grader per task and human baselines, compatible with AISI's Inspect platform (SpecOps Inc for the UK AI Security Institute, WP 1) $1.3M, cyber range
- Cyber range scenarios for critical national infrastructure built to order for AI evaluation (Hack The Box for the UK AI Security Institute, Cyber Evaluations WP1) $782K, cyber range
- Non-public capture-the-flag scenarios and end-to-end attack exercises, plus a benchmark suite of scenarios testing whether LLM agents' safeguards can be bypassed (Grey Swan Security for the UK AI Safety Institute) $583K, cyber ctf
- Suites of tasks ('evals') for evaluating AI systems, created to order (Rethink Priorities for the UK AI Safety Institute) $251K, eval tasks
Made to orderper task
$2,837
Middle half $826 to $5,972. 8 figures (+2 open-ended), 16 Dec 2023 to 15 Oct 2026.
The figures
- Legal-environment RL tasks ordered at scale from Mercor (the most common kind of order) $2,000, legal
- RL training tasks sold by Mercor to AI labs, full price range $707
- Tasks built to a buyer's request on HUD's DataVendor marketplace $866
- One decent RL coding task bought by a frontier lab from a data or environment company $5,000, coding
- One especially complex software-engineering RL task with its grader (upper end reported to Epoch AI) $20K, coding
- One training task with its grader $632
- Recommended spend per RL task for frontier post-training (Mechanize analysis) ~$500 per RL task now, rising roughly fivefold to a few thousand dollars within a year
- One successful dangerous-capability or autonomous-agent evaluation task built for the UK AI Safety Institute (evals bounty) $8,886
- Private agent-evaluation task for METR, priced by the hours a professional would need to complete it $300 per hour it would take a professional to complete the task, coding
- One private agent-evaluation task implementation for METR that takes an experienced engineer about 10 hours (worked example) $3,674, coding
Made to orderper environment
$42.7K
Middle half $28.3K to $56.9K. 4 figures, 12 Jan 2026 to 16 Oct 2026.
The figures
- Custom sim2real digital twin RL environment: calibrated dynamics, domain randomisation, validation on real hardware in the vendor's San Francisco lab, transfer notebook, 6 to 8 weeks $50K, robotics sim
- Custom simulated robotics RL environment, Starter tier: 1 robot, 1 simulator (Isaac Sim, MuJoCo, Genesis or Gazebo), 1 to 3 tasks, Gymnasium API, reward functions, source code, one training-ready dataset, 2 weeks $7,071, robotics sim
- Custom simulated robotics RL environment suite: multi-robot, multi-simulator, 5 to 10 tasks, calibrated dynamics, domain randomisation, VLA-ready evaluation harness, 4 to 6 weeks $35.4K, robotics sim
- A working replica of a website or app for an agent to practise in $77.5K
Made to orderper hour of work
$187too few figures for a median
1 figure, 5 Nov 2024.
Cost to produceper demonstration
$0.42
Middle half $0.093 to $0.74. 6 figures, 27 Nov 2024 to 4 Oct 2025.
The figures
- Human computer-use demonstrations across 3 operating systems and 200+ apps (OpenCUA AgentNet, 22K tasks) $0.91
- Synthetic multimodal web-agent trajectory from exploration (Explorer, 94K successful trajectories) $0.28
- 150,000 synthetic web-agent trajectories with judge annotations, generated on 1,200 V100 GPU hours (InSTA) $0.0035
- Synthetic GUI/web agent trajectory generated by replaying web tutorials (AgentTrek) $0.55
- Human-annotated web-agent trajectory in the Mind2Web training set (as costed by the Synatra authors) $0.8
- Synthetic web-agent demonstration converted from tutorials and indirect knowledge (Synatra) $0.031
Cost to produceper task
$1,500
Middle half $400 to $2,400. 5 figures, 7 Apr 2025 to 15 Oct 2026.
The figures
- Cost basis of one expert-written RL task at Mercor (ten hours of expert time at $150 an hour) $1,500
- Labour cost of one good RL coding task at Mechanize: an engineer paid $400k+ a year expected to make one good task a week $7,692, coding
- Synthetic chain-of-thought reasoning added to each human computer-use demonstration (OpenCUA) $0.6
- Lifetime compute for one frontier RL task (500,000 tokens per run, group size 64, five reuses, at Grok 4 output-token prices; Mechanize analysis) $2,400
- One 12-hour agent rollout replicating a research paper (o1 IterativeAgent on PaperBench) $400
Cost to produceper dataset
$500Ktoo few figures for a median
1 figure, 15 Oct 2026.
By segment, where two or more figures exist
- Codingpaid to the people who make it, per environment$3,2504 figures
- Ml engineeringpaid to the people who make it, per environment$1,200 to $2,0002 figures
- Computer usepaid to the people who make it, per environment$1,7506 figures
- App simulatorpaid to the people who make it, per environment$1,5009 figures
- Tool usepaid to the people who make it, per environment$500 to $1,5002 figures
- Scientific computingpaid to the people who make it, per task$200 to $3002 figures
- Computational biologypaid to the people who make it, per task$150 to $2602 figures
- Eval tasksmade to order, per contract$251K to $3M2 figures
- Cyber rangemade to order, per contract$782K to $1.3M2 figures
- Cyber ctfmade to order, per contract$583K to $722K2 figures
- Codingmade to order, per task$5,0003 figures
- Robotics simmade to order, per environment$35.4K3 figures
Physical-world data
First-person video of manual work, teleoperation and motion capture, for robots.
Paid to the people who make itper hour of work
$20
Middle half $8.37 to $32. 33 figures (+1 open-ended), 14 May 2024 to 11 Oct 2026.
The figures
- Data collection operator using UMI, GELLO and VR teleoperation tools (RoboForce, Milpitas, swing shift) $32, teleoperation
- Proposed common hourly rate for people who collect robot-training video in India (industry body) $8.37, egocentric
- Mobile egocentric video dataset, chest or head-mounted phone, minimum 60 clips totalling 90 minutes, India $10, egocentric
- Video of repeated hand manipulation tasks with handheld devices for robotics, United States $14, manipulation video
- Contractor performing repetitive manipulation tasks with handheld devices for robot dexterity data (micro1 for a client, Los Angeles) $14, umi
- Senior software engineer running a teleoperation data-collection floor and directing its operators (Welo Data, Project Hydra, Santa Clara) $105, teleoperation
- Operator running a handheld UMI gripper through repetitive manipulation tasks in a San Francisco lab (micro1 for a client) $30, umi
- Data collection operator (swing shift) generating real-world training data for the NEO home humanoid at 1X (San Carlos) $32, teleoperation
- Egocentric 'day in a life' recordings on a head-mounted Insta360 X5 in New Zealand (RemoExperts) $110, new zealand
- Egocentric 'day in a life' recordings on a head-mounted Insta360 X5 in Singapore (RemoExperts) $120, singapore
- QC reviewer of robot manipulation video annotations (Welo Data, Project Cursa, Philippines) $4.03, philippines
- Annotator writing three-level natural-language labels for three-camera robot manipulation videos (Welo Data, Project Cursa, Philippines) $3.50, philippines
- At-home household video recording, United States (entry level) $12, egocentric
- Robot operator providing teleoperated demonstrations at Physical Intelligence (San Francisco) $25, teleoperation
- Robot operator doing VR teleoperation for Toyota Research Institute (via HireArt) $35, teleoperation
- In-house collector recording task demonstrations and testing new tasks and hardware (Sunday Robotics, Redwood City, evening shift) $35, glove demonstration
- Indian garment worker wearing a head camera through a factory shift about Rs 400 daily, india
- On-site collector wearing a first-person camera while doing household activities (Appen/CrowdGen, Kuala Lumpur, 4-day project) $3.60, malaysia
- Recording short videos of everyday household or professional tasks with a recording headstrap for robot training (MicroAGI, Shift app) $20, egocentric capture
- Egocentric (first-person) video collection by workers in India for Human Archive $1.00, egocentric capture
- Egocentric data collection workers in India paid by Human Archive's competitors $3.32, india
- Data collector teleoperating robot arms through diverse tasks at Dyna Robotics (Redwood City) $31, teleoperation
- Hourly wages for human data collection for robot training (people recording everyday tasks), industry estimate by region $10, egocentric capture
- At-home recording of household tasks for humanoid robot training (micro1, Nigeria) $15, egocentric capture
- Operator collecting humanoid robot data in VR (1X, listed under Kind Humanoid, San Carlos, day shift) $32, teleoperation
- Senior operator collecting humanoid robot data in VR, in lab and home settings (1X, listed under Kind Humanoid, San Carlos) $41, teleoperation
- Robot operator doing teleoperation and data collection at Skild AI (Pittsburgh) $19, teleoperation
- Egocentric POV video of household tasks for humanoid training datasets, Philippines $6.71, egocentric
- Labelling autonomous-vehicle driving data (validating AI pre-labels on edge cases) $4.24, driving
- Remote worker filming their own household chores first-person with a head-mounted phone for robotics (micro1) $8.00, egocentric
- Robot operator doing teleoperation and data collection at Skild AI (San Mateo) $27, teleoperation
- Skilled participant recorded in first- and third-person video for the Ego-Exo4D dataset $20, egocentric
- Data collection operator wearing a motion-capture suit and VR headset for Tesla's Optimus program (Palo Alto) $35, motion capture
- Humanoid robot pilot wearing teleoperation equipment to guide Figure's robot and upload the data (San Jose) $32, teleoperation
Paid to the people who make itper hour of video
$14
Middle half $6.00 to $17. 14 figures (+1 open-ended), 19 Nov 2025 to 7 Oct 2026.
The figures
- Car-mounted 360-degree exterior driving video (trailing third-person view of turns, stops, intersections, parking), recorded by contributors in 52 countries $17, driving
- Driver recording their own car's road navigation from a trailing 360-degree camera (micro1 client project) $17, driving
- Third-person 360-degree video of movement and navigation (Insta360 or GoPro MAX) for a micro1 client $20, exocentric
- EgoData egocentric POV video of household activities, iPhone 12+, India $2.27, egocentric
- EgoData egocentric POV video in commercial or workplace settings, iPhone 12+, India $3.31, egocentric
- At-home recorder filming household chores first-person on a smartphone headstrap (micro1) $15, egocentric
- At-home first-person video of household tasks, United States $14, egocentric
- At-home first-person video of repetitive physical movements, Argentina, Brazil, Chile, Colombia, Mexico $6.00, egocentric
- First-person household video data collection, United States (participating states) $14, egocentric
- First-person household chores video capture, Latin America $6.00, egocentric
- First-person garage and workshop task video capture, India, Philippines, Indonesia, Vietnam $5.00, egocentric
- First-person garage and workshop task video capture, Canada $10, egocentric
- First-person garage and workshop task video capture, United States $15, egocentric
- Remote collector recording more complex household-task demonstrations after screening (Sunday Robotics) Up to $60 per hour of approved data, glove demonstration
- Remote collector recording household-task demonstrations on provided hardware (Sunday Robotics, screening week) $30, glove demonstration
Sales that happenedper hour of video
$0.58 to $2.59too few figures for a median
2 figures, 10 Oct 2026.
Asking prices off the shelfper hour of video
$8.88
Middle half $4.48 to $25. 16 figures (+3 open-ended), 29 Oct 2025 to 11 Oct 2026.
The figures
- Unidata Egocentric Video Dataset, 4,050 hours of first-person household video with 3D pose and IMU (Pico + Zed + motion trackers) Starts at $30,000 / purchase, egocentric
- Unidata Egocentric Video Dataset, Pico + Motion Trackers setup, 2,321 hours of first-person household manipulation (13 scenarios) with IMU Starts at $30,000 / purchase, egocentric
- Unidata Egocentric Video Dataset, Zed stereo + Pico + motion trackers setup, 1,729 hours of multimodal first-person household recordings Starts at $30,000 / purchase, egocentric
- Infobay.AI 5K+ Hours Stereo IMU Egocentric Dataset, stereo video with synchronized IMU $27, egocentric
- Shanghai Zhixiangyuan China Egocentric Human Activity Dataset, ~3,000 hours residential and commercial first-person video, mostly 1080p ~60 fps $2.85, egocentric
- Infobay.AI 220K+ Hours Narrated Egocentric Video Dataset, first-person video with embedded spoken narration (price for initial 50,000-hour batch) $9.76, egocentric
- Verbose TechLabs Egocentric Factory Assembly Line, USA, 25,000 hours of head/chest-mounted video $4.47, egocentric
- Verbose TechLabs Egocentric Household Maintenance, Colombia, 250 hours (cleaning, plumbing, furniture) with IMU $4.48, egocentric
- Verbose TechLabs Egocentric Living Room & Bedroom, USA, 34 hours of casual indoor first-person video $4.71, egocentric
- Verbose TechLabs Egocentric Precision Electronics, Japan, 10,000 hours of cleanroom/microelectronics first-person video $5.00, egocentric
- Deploy8 50-Hour Industrial Egocentric Stereo, electric motor rewinding workshops in India, 49.3 hours, 117 sessions, raw $8.01, egocentric
- Nexdata 10,000-Hour Egocentric Full-Body Multimodal Dataset, 4K stereo (PICO 4 Ultra) with 76-point full-body pose and step labels $66, egocentric
- Merit Data South Asia Residential Mono Egocentric Video Dataset, ~800 hours / ~10,000 iPhone recordings, 4K and 1080p $79, egocentric
- Infobay.AI 38K+ Hours IMU Egocentric Video Dataset, wearable first-person video with accelerometer and gyroscope streams $15, egocentric
- First-person video of manual work, in bulk $3.16, egocentric
- Annotated Indian egocentric footage resold to global robotics labs $27, india
- Raw egocentric video hours from India, priced by task complexity (Human Archive) $3.16, egocentric
- First-person video of manual work, captured to a spec $24, egocentric
- JoinAI POV Robotic Behavior Video Dataset, 1,000,000+ generated first-person 6 to 8 s clips, 2,000+ hours, 1080p $20, synthetic
Asking prices off the shelfper record
$0.66
Middle half $0.19 to $1.17. 6 figures, 17 Mar 2026 to 14 Oct 2026.
The figures
- AI-detected driving event clip with video and sensor data (Bee Maps AI Event Videos API) $6.00, driving
- Machine-detected road features (signs, lane lines and the like) from dashcam imagery (Bee Maps Road Detections) $0.003, driving
- MoCap Online crowd animation pack, 108 optical motion-capture animations $0.32, motion capture
- MoCap Online ladder animation pack, 198 optical motion-capture animations $0.15, motion capture
- MoCap Online walk animation pack, 30 optical motion-capture animations $1.00, motion capture
- One cleaned motion-capture animation bought in a pre-made pack $1.22, motion capture
Asking prices off the shelfper image
$0.005too few figures for a median
1 figure, 14 Oct 2026.
Asking prices off the shelfper dataset
$2,850too few figures for a median
1 figure (+5 open-ended), 11 Oct 2026 to 14 Oct 2026.
The figures
- Bimanual Manipulation Pack, ready-to-license robot dataset (Robotics Center) Bimanual Manipulation Pack: From $5,000, teleoperation
- Dexterous Grasping Pack, ready-to-license robot dataset (Robotics Center) Dexterous Grasping Pack: From $5,000, teleoperation
- Humanoid Locomotion Pack, ready-to-license robot dataset (Robotics Center) Humanoid Locomotion Pack: From $10,000, teleoperation
- VLA Evaluation Pack, ready-to-license robot dataset (Robotics Center) VLA Evaluation Pack: From $3,000, teleoperation
- Robot 'Mirror Suit' robotics and vision imagery, 425 assets (BWS on Datarade) $2,850
- Real-world egocentric POV human task video dataset for physical AI and robotics (M-ART on Datarade) Starts at $10,000 / purchase, egocentric
Asking prices off the shelfper demonstration
$101too few figures for a median
1 figure, 29 Jul 2026.
Made to orderper hour of video
$38
Middle half $15 to $59. 20 figures, 18 May 2026 to 24 Sep 2026.
The figures
- First-person (egocentric) human data collected in China, quoted price per hour $15, egocentric
- Real-robot teleoperation data collected in China, quoted price per hour $106, teleoperation
- Egocentric-only household task video capture $126, egocentric
- Synchronized multi-camera and multi-sensor robotics video capture $300, multi sensor
- Human annotation of egocentric robotics video (Macrodata benchmark, cited by Perceptron) $50, egocentric
- Bimanual ALOHA-style teleoperation program, managed by DataXpower $57, teleoperation
- Egocentric wearable program with annotation (head and wrist cameras), managed by DataXpower $39, egocentric
- Full humanoid multi-sensor program (mocap suits or exoskeletons, F/T and IMU), managed by DataXpower $110, motion capture
- Simple single-arm teleoperation program (pick-and-place, sorting), managed by DataXpower $21, teleoperation
- Worked budget for a 10,000-hour teleoperation dataset with language annotation for a VLA model (DexSet) $65, teleoperation
- Annotation passes on robot data, added to any capture modality (DexSet benchmark) $14
- Dense object masks or contact annotation on robot data (DexSet per-pass price) $21
- Language-instruction labelling of robot data (DexSet per-pass price) $9.80
- Multi-view exocentric capture for robot learning (DexSet benchmark) $32, exocentric
- Teleoperation demonstrations for robots $41, teleoperation
- Payment from egocentric-data startups to an Indian factory whose workers wear cameras (RoBoEra figure) $5.02, india
- Crowd-platform (unmanaged) video collection in India $3.46, crowdsourced
- Managed indoor egocentric video collection in the UK and EU $44, egocentric
- Managed indoor egocentric video collection from US vendors, comparable quality $37, egocentric
- Managed indoor egocentric (first-person) video collection in Vietnam $15, egocentric
Made to orderper demonstration
$13
Middle half $7.42 to $70. 14 figures, 31 Mar 2026 to 14 Oct 2026.
The figures
- Humanoid or dexterous full-body teleoperation episode, whole-body policy training (Robotics Center custom collection) $8.66, teleoperation
- Long-horizon, bimanual or contact-rich teleoperation episode, production policy fine-tuning (Robotics Center custom collection) $3.87, teleoperation
- Mid-complexity manipulation teleoperation episode (15 to 30 seconds), standard imitation learning (Robotics Center custom collection) $1.73, teleoperation
- Teleoperation task episode on a novel embodiment (robot shipped in by the customer), bespoke R&D collection (Robotics Center) $500, teleoperation
- Simple pick-and-place teleoperation episode (about 5 seconds), for VLA pretraining volume (Robotics Center custom collection) $0.71, teleoperation
- 2,000-demonstration bimanual imitation-learning dataset for one task family (DataXpower planning figure) $77, teleoperation
- Collection of 5,000 whole-body humanoid demonstrations (one program) $87
- Complex humanoid whole-body demonstration (DataXpower benchmark) $87
- Contact-rich bimanual robot demonstration episode (Robotics Center cost guide) $30, teleoperation
- Simple pick-and-place robot demonstration episode (Robotics Center cost guide) $11, teleoperation
- Managed teleoperation collection of 2,000 demonstrations (Robotics Center) $10, teleoperation
- Managed teleoperation collection of 50 demonstrations (Robotics Center) $50, teleoperation
- Managed teleoperation collection of 500 demonstrations (Robotics Center) $16, teleoperation
- Managed teleoperation collection of 5,000 demonstrations (Robotics Center) $7.00, teleoperation
Made to orderper hour of work
$94
Middle half $75 to $102. 4 figures, Mar 2026 to 11 Oct 2026.
The figures
- Robot teleoperation data (Claru's market figure) $100, teleoperation
- Skilled motion-capture editor time for cleaning raw capture $106, motion capture
- Robot teleoperator time bought from operators in India, the Philippines and Eastern Europe $35, teleoperation
- Robot teleoperator time bought from US-based operators with domain expertise $88, teleoperation
Made to orderper contract
$2,121
Middle half $1,458 to $51.1K. 3 figures (+1 open-ended), 17 Mar 2026 to 14 Oct 2026.
The figures
- Full day in Teesside University's VR motion-capture studio (18 infra-red cameras, up to six suits), motion-capture control with technical support $795, motion capture
- Embodied AI data collection and annotation service for robots (Nexdata on Datarade) Starts at $10,000 / purchase
- Production-grade robot training dataset of 500 or more demonstrations (Robotics Center cost guide) $100K, teleoperation
- One day of a professional mid-tier optical motion-capture studio (OptiTrack or Vicon) $2,121, motion capture
Made to orderper dataset
$1,000too few figures for a median
1 figure, 11 Oct 2026.
The figures
- Egocentric video sample pack, 10 to 15 clips with basic metadata $1,000, egocentric
Made to orderper record
$100too few figures for a median
1 figure, 17 Mar 2026.
The figures
- One motion-capture clip from a professional optical studio session, after cleanup $100, motion capture
Cost to produceper hour of video
$127
Middle half $93 to $187. 4 figures, Mar 2024 to 11 Oct 2026.
The figures
- Egocentric video from people wearing cameras in everyday activities (Claru's cost figure) $17, egocentric
- Fully loaded cost of one hour of labelled, RLDS-formatted robot manipulation teleoperation data (Robotics Center index) $118, teleoperation
- Fully loaded cost of one hour of labelled, RLDS-formatted robot manipulation teleoperation data (Robotics Center index) $136, teleoperation
- Fully loaded cost of one hour of labelled, RLDS-formatted robot manipulation teleoperation data (Robotics Center index) $340, teleoperation
Cost to produceper demonstration
$6.88
Middle half $2.96 to $13. 4 figures, 31 Mar 2026 to 27 May 2026.
The figures
- Real-robot teleoperation data in China, cost per task episode from equipment amortisation and sundries $0.58, teleoperation
- Production cost of one usable bimanual assembly demonstration, all components (Robotics Center cost model) $10, teleoperation
- Production cost of one usable dexterous manipulation demonstration, all components (Robotics Center cost model) $20, teleoperation
- Production cost of one usable simple pick-place demonstration, all components (Robotics Center cost model) $3.75, teleoperation
Cost to produceper hour of work
$16 to $26too few figures for a median
2 figures, 2 Jul 2026.
The figures
Cost to produceper record
$16too few figures for a median
1 figure, 17 Mar 2026.
The figures
By segment, where two or more figures exist
- Teleoperationpaid to the people who make it, per hour of work$3210 figures
- Umipaid to the people who make it, per hour of work$14 to $302 figures
- Egocentric capturepaid to the people who make it, per hour of work$134 figures
- Egocentricpaid to the people who make it, per hour of work$9.196 figures
- Philippinespaid to the people who make it, per hour of work$3.50 to $4.032 figures
- Egocentricpaid to the people who make it, per hour of video$8.0010 figures
- Egocentricsales that happened, per hour of video$0.58 to $2.592 figures
- Egocentricasking prices off the shelf, per hour of video$6.5014 figures
- Drivingasking prices off the shelf, per record$0.003 to $6.002 figures
- Motion captureasking prices off the shelf, per record$0.664 figures
- Teleoperationmade to order, per hour of video$575 figures
- Egocentricmade to order, per hour of video$397 figures
- Teleoperationmade to order, per demonstration$1012 figures
- Teleoperationmade to order, per hour of work$883 figures
- Motion capturemade to order, per contract$795 to $2,1212 figures
- Teleoperationcost to produce, per hour of video$1363 figures
- Teleoperationcost to produce, per demonstration$6.884 figures
Work by people
Records made to order: an expert's hour, a recorded workflow, a document its author owns.
Paid to the people who make itper hour of work
$77
Middle half $55 to $101. 105 figures (+6 open-ended), 18 Jan 2023 to 17 Oct 2026.
The figures
- Median advertised hourly pay for Agentic Tasks & RL Environments roles in AI training $50, agentic tasks rl environments
- Median advertised hourly pay for AI Evaluation & Rating (RLHF) roles in AI training $60, ai evaluation rating rlhf
- Median advertised hourly pay for Coding & Software Evaluation roles in AI training $65, coding software evaluation
- Median advertised hourly pay for Data Annotation & Labeling roles in AI training $30, data annotation labeling
- Median advertised hourly pay for Law, Medicine & Finance roles in AI training $70, law medicine finance
- Median advertised hourly pay for Math & STEM roles in AI training $65, math stem
- Median advertised hourly pay for Multilingual & Localization roles in AI training $65, multilingual localization
- Median advertised hourly pay for Red-Teaming & AI Safety roles in AI training $55, red teaming ai safety
- Median advertised hourly pay for Research Studies & Surveys roles in AI training $100, research studies surveys
- Median advertised hourly pay for Robotics & Physical-AI Data roles in AI training $10, robotics physical ai data
- Median advertised hourly pay for Transcription, Audio & Speech roles in AI training $30, transcription audio speech
- Median advertised hourly pay for Writing & Content roles in AI training $55, writing content
- Biologist, PhD at Alignerr (PhD biologists writing and reviewing advanced AI evaluation questions, posted for Labelbox) $104, biology
- Biology Masters at Alignerr (Master's-level biologists evaluating AI answers) $44, biology
- Chemist, PhD at Alignerr (PhD chemists writing and reviewing advanced AI evaluation questions, posted for Labelbox) $104, chemistry
- Chemistry Masters at Alignerr (Master's-level chemists evaluating AI answers) $44, chemistry
- Mercor's average contracted expert rate (homepage) $112, all
- Lawyers (Mercor listing) $110, law
- Chemistry Specialist, Freelance AI Trainer Project (Meridial, global remote contractor) $23, chemistry
- All AI-training work: live median advertised hourly pay across 26 platforms (AITraining.jobs pay report) $65, all
- Handshake AI: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $90, all
- Terac: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $85, all
- Dutch Audio Transcriptionist (AI Training) at Alignerr $19, languages
- Medical expert work on Mercor, nurses and other clinical professions (standing talent network band) $104, medicine
- Physician (MD/DO) work on Mercor, clinical documentation and AI evaluation (standing talent network band) $131, medicine
- Babel Audio: median advertised hourly pay across its live roles $105, voice
- Mindrift (Toloka): median advertised hourly pay across its open AI-training roles (AITraining.jobs) $100, all
- RemoExperts: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $95, all
- Sovrano AI: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $80, all
- Mercor: Consulting Expert (role on the homepage) $95, generalist
- Mercor: Legal Expert (role on the homepage) $90, law
- AI Tutor roles in 13 languages at xAI (Croatian, Finnish, French, Hausa, Italian, Khmer, Korean, Norwegian, Portuguese, Serbian, Spanish, Swedish, Yoruba) $40, languages
- BioInformatics Experts at Mercor (writing multi-step bioinformatics data-analysis tasks with validated answers) $99, biology
- AfterQuery: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $75, all
- Babel Audio: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $105, all
- Mercor: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $80, all
- micro1: median advertised hourly pay across its open AI-training roles (AITraining.jobs) $70, all
- Accounting & Finance Expert at Alignerr $77, finance
- AI Red Team Tester at Alignerr $69, red teaming
- Average pay across Alignerr (Labelbox) experts $80, all
- Biology Expert (Masters/PhDs) at Alignerr $82, biology
- Chemistry Expert (Masters/PhDs) at Alignerr $82, chemistry
- Creative Writing Evaluator at Alignerr $69, writing
- Lean 4 Proof Engineer - Mathematical Formalization (Alignerr) $184, math
- Legal & Government Policy Analyst (AI Training) at Alignerr $28, law
- Mathematics Expert (Masters/PhDs) at Alignerr $82, math
- Machine Learning Evaluation Specialist at Alignerr (designing ML challenges that stump frontier models) $283, coding
- Oncology Clinical Researcher at Alignerr $57, medicine
- Physicist, PhD at Alignerr $104, physics
- Software Developer paid to build their own project with an AI coding assistant (Alignerr) $85, coding
- Voice Actor - German Speakers at Alignerr $265, voice
- Software engineering, data science and ML roles on DataAnnotation $77, coding
- Finance, legal, medical, cybersecurity, engineering and STEM roles on DataAnnotation $71, professional
- Handshake AI experts in mathematics, physics and computer science $112, expert
- AI Evaluation Specialist (Handshake AI fellowship, trending opportunity) Up to $17/hr, generalist
- FP&A Analyst (Handshake AI fellowship, trending opportunity) Up to $80/hr, finance
- Philosophy Expert (Handshake AI fellowship, trending opportunity) Up to $120/hr, humanities
- Software Engineer (Handshake AI fellowship, trending opportunity) Up to $120/hr, coding
- Applied Physics Benchmark Specialist role on Mercor's live board $69, physics
- Data Science Expert role on Mercor's live board $122, data science
- Inorganic Chemistry & Synthesis Expert (PhD-level evaluation of materials synthesis problems) at Mercor $84, chemistry
- Mindrift (Toloka) AI trainers, domain experts $77, law medicine finance
- Mindrift (Toloka) AI trainers, entry level $21, generalist
- A software expert recording and narrating real work on screen (Mercor listing) Up to $65 an hour, workflow recording
- Prolific participant and tasker pay, minimum $8.00, generalist
- Prolific participant and tasker pay, recommended $12, generalist
- Surge AI expert contractor (award-winning author or journalist tier) $283, writing
- Oncologist (Mercor listing) $397, medicine
- Research Advisor, Human Frontier Collective (Scale AI): senior finance, legal and medical experts $300, law medicine finance
- AI Tutor, Video at xAI $55, video
- AI Red Teamer (LLM generalist, Remote) at Handshake AI $55, red teaming
- AI Tutor, Design at xAI $51, design
- AI Tutor, Humanities at xAI $51, humanities
- Lean Engineer, formal mathematics (Lean 4, Mathlib) (Mercor) $99, math
- AI Tutor, Legal and Compliance at xAI $71, law
- AI Tutor, Software Engineering at xAI $63, coding
- AI Tutor, Writing at xAI $51, writing
- Condensed Matter Physicist (PhD) (Mercor) $94, physics
- Biotechnology Expert talent network at Mercor (bioprocess and therapeutic judgment for AI model evaluation) $117, biology
- Chemistry Expert talent network at Mercor (graduate-level chemistry judgment for AI model evaluation) $83, chemistry
- Audiobook QA Expert, English (US) (Mercor) $22, voice
- Equity Research Expert, single-name analyst publications (Mercor) $194, finance
- Application user recording generalist software use (Photoshop on Windows) (Mercor) $35, generalist
- Molecular Biology Experts, US (Mercor) $86, biology
- AI Red Teamer, CBRNE (Remote) at Handshake AI $101, red teaming
- AI Red Teamer, Cybersecurity (Remote) at Handshake AI $90, red teaming
- AI Model Policy Trainer, Image Evaluation (Seattle onsite) at Handshake AI $51, evaluation
- AI Safety Policy Evaluator, Violence and Threats (Seattle onsite) at Handshake AI $81, red teaming
- Drug Discovery Scientist at Mercor (building drug R&D evaluation problems for frontier AI models) $73, biology
- AI Model Policy Trainer, Generalist (Seattle) at Handshake AI $50, generalist
- Spanish Tutors (Mercor) $62, languages
- AI Model Policy Trainer, Mental Health (Seattle onsite) at Handshake AI $54, medicine
- Clinicians (Mercor listing) $134, medicine
- Biochemistry & Life Sciences Domain Expert at Mercor, embedded with a leading AI lab's GenAI team (USA only) $83, biology
- Investment Banking & M&A finance domain expert (Mercor) $122, finance
- Partner & General Counsel, legal domain expert (Mercor) $122, law
- AI Safety Expert, English and Norwegian (Mercor) $55, red teaming
- AI Safety Expert, English and Vietnamese (Mercor) $21, red teaming
- Mercor's average pay to about 30,000 contractors (H1 2026) $105, all
- AI Safety Red Teamer (Mercor) $77, red teaming
- Mercor's advertised headline rate (early June 2026) $141, all
- Writing expert role on Mercor $106, writing
- Voice Actor for CX agent voice cloning, USA (Mercor) $87, voice
- Chemist Talent Network at Mercor (open application for future AI-training contracts in chemistry) $69, chemistry
- Effective pay for a generalist tasker on some Outlier (Scale AI) projects $20, generalist
- Generalist data work on Mercor's Meta project Nova, after project Musen ended $16, generalist
- Mercor's top industry-expert rate (October 2025) up to $200 an hour, law medicine finance
- Former investment bankers writing prompts and building financial models for OpenAI (Project Mercury) $150, finance
- Mercor's average pay for experts (October 2025) $85, all
- Annotator writing IMO-level math problems for RLHF (Abaka AI cost model) $20, math
- Sama data labellers in Kenya working on OpenAI's ChatGPT safety project $1.62, red teaming
Paid to the people who make itper task
$212
Middle half $167 to $718. 4 figures (+2 open-ended), 8 Aug 2024 to 11 Oct 2026.
The figures
- Chemists, explosives and nuclear security specialists writing safety tasks on Mercor $70, chemistry
- A developer's own AI coding session, shared as data (Mercor) $200, coding
- Database Systems Evaluation Expert, per task (Mercor) $2,200, coding
- Agent Engineer, per task (Mercor) $224, coding
- A verified universal jailbreak of Anthropic's unreleased safety classifiers (bug bounty) up to $25,000, red teaming
- A verified universal jailbreak of Anthropic's unreleased safety classifiers (bug bounty) up to $15,000, red teaming
Paid to the people who make itper contract
Open-ended only
0 figures (+5 open-ended), 21 Nov 2025 to 11 Oct 2026.
The figures
- Alignerr (Labelbox) weekly payouts to experts $1M Paid weekly, all
- Cumulative payouts to Handshake AI fellows $500M+ payouts made, all
- Mercor daily payouts to experts $4M+ Daily payouts, all
- Cumulative payouts to Outlier (Scale AI) contributors $500M+ Paid to experts, all
- Handshake AI monthly payouts to its experts, against its gross run rate ~$60M a month to over 30,000 experts; ~$1B run rate, all
Offered to holdersper document or page
$6.00too few figures for a median
1 figure, 1 Aug 2026.
Made to orderper task
$0.08
Middle half $0.032 to $0.22. 17 figures, 17 Apr 2026 to 17 Oct 2026.
The figures
- CVAT annotation services, one-time project: per annotated object, from its worked example (minimum project $5,000) $0.1, annotation
- CVAT annotation services, 6-month prepaid subscription (from $5,000): reduced per-object rate $0.061, annotation
- Label Your Data: bounding box annotation, starting price $0.02, annotation
- Label Your Data: keypoint annotation, starting price $0.015, annotation
- Label Your Data: NLP annotation, starting price per entity $0.02, annotation
- Precise BPO: bounding box annotation, per-task rate range (dedicated offshore team) $0.024, annotation
- Roboflow Outsource Labeling: bounding box annotation, starting rate $0.1, annotation
- Roboflow Outsource Labeling: classification or keypoint annotation, starting rate $0.05, annotation
- Roboflow Outsource Labeling: polygon annotation, starting rate $0.2, annotation
- US-DATA: image bounding box annotation, indicative price per object $0.08, annotation
- US-DATA: named entity recognition (NER) text annotation, indicative price per text $0.25, annotation
- Simple annotation on Appen's self-service platform $0.032, annotation
- Annotation task bought from Appen, typical per-task price across complexity $0.071, annotation
- Annotation task bought from Scale AI, typical per-task price across complexity $0.22, annotation
- Expert RLHF task such as grading code or medical answers, bought from an annotation vendor $45, code medical grading
- Nuanced RLHF evaluation requiring domain expertise (one comparison set) $45, expert evaluation
- Domain-expert RLHF annotation (medical, legal, financial) $3.46, law medicine finance
Made to orderper hour of work
$19
Middle half $13 to $94. 9 figures, 18 Jan 2023 to 14 Oct 2026.
The figures
- Data annotation billed by time (Label Your Data) $6.00, annotation
- Clinical Annotator III: GSA schedule ceiling rate billed to US government buyers (FedPoint Systems) $125, medicine
- Data Labeler III (TS/SCI): GSA schedule ceiling rate billed to US government buyers (Chenega Technology & Modernized Solutions) $73, labeling
- Data Labeler, Intermediate: GSA schedule ceiling rate billed to US government buyers (Chenega Systems) $94, labeling
- Medical Annotator III: GSA schedule ceiling rate billed to US government buyers (FedPoint Systems) $153, medicine
- Prolific platform fee on participant and AI-tasker rewards (academic or non-profit customers) $13, generalist
- Prolific platform fee on participant and AI-tasker rewards (corporate customers) $14, generalist
- Data annotation billed by the annotator hour, all regions and skill levels $19, annotation
- Text labelling for ChatGPT safety, outsourced by OpenAI to Sama in Kenya (late 2021 contracts) $13, red teaming
Made to orderper record
$2.58
Middle half $1.25 to $5.29. 6 figures, 9 Sep 2024 to 5 Aug 2026.
The figures
- Generalist RLHF preference ranking (one pairwise comparison), bought from an annotation vendor $2.00, preference ranking
- RLHF preference label for reward modelling, crowdsourced $1.00, preference ranking
- RLHF preference label for reward modelling, by domain experts $7.07, preference ranking
- Simple RLHF preference judgment on a straightforward query (one comparison) $3.16, preference ranking
- General RLHF preference comparison (one pair judged) $0.22, evaluation
- Supervised fine-tuning (SFT) data, one constructed entry $6.00, writing
Made to orderper contract year
$150M to $200Mtoo few figures for a median
2 figures (+3 open-ended), 14 Jun 2025 to 1 Sep 2026.
The figures
- Google's annual contract with Surge AI more than $100 million per year, all
- Meta's 2024 spend with Surge AI $150M, all
- micro1 net annual run rate after paying experts, on a $500M gross run rate net annual run rate between $150 million and $200 million, all
- Handshake AI's largest frontier AI lab customer accounts $100M+ run rate per account, all
- Google's planned 2025 spend with Scale AI for human-labeled training data $200M, all
Made to orderper contract
$9.95too few figures for a median
1 figure (+13 open-ended), 18 Jan 2023 to 11 Oct 2026.
The figures
- Handshake AI data business gross margin gross margins around ~30%, all
- Share of Mercor's gross revenue paid to contractors 60-70% of top-line revenue, all
- OpenTrain AI managed AI-training teams Cost-plus 20% management fee, all
- OpenTrain AI self-service fee on payments to AI trainers and annotators $9.95, all
- Surge AI's prices relative to competing human-data vendors 50% to ten times more than competitors, all
- Mercor gross margin on AI-lab billings, 2025 27% (2025), all
- Mercor gross margin on AI-lab billings, Q2 2026 33% (Q2 2026), all
- Mercor gross margin on AI-lab billings, 2027 target 46% (2027 target), all
- Mercor gross revenue in the first half of 2026, and the share of it from AI foundation model companies $614 million, more than 90% from AI foundation model companies, all
- Share of every dollar kept by data vendors that run delivery from lower-cost geographies 70-80%+, all
- Share of Mercor's topline revenue paid out to its contractors 60% to 70%, all
- Enterprise LLM alignment (RLHF) programme, total $50,000 to several hundred thousand dollars, evaluation
- Mercor markup on contractor billing (February 2024) approximately 35%, all
- Three OpenAI contracts with Sama for labelling work about $200,000 in total, red teaming
Made to orderper demonstration
$12too few figures for a median
1 figure, 28 Apr 2026.
The figures
- Human-written SFT demonstration dataset of 10,000 demonstrations $12, sft demonstration
Made to orderper dataset
Open-ended only
0 figures (+1 open-ended), 20 Aug 2026.
The figures
- micro1 gross margin on off-the-shelf datasets resold to several customers as high as 80% to 90%, all
By segment, where two or more figures exist
- Financepaid to the people who make it, per hour of work$1364 figures
- Medicinepaid to the people who make it, per hour of work$1186 figures
- Voicepaid to the people who make it, per hour of work$964 figures
- Physicspaid to the people who make it, per hour of work$943 figures
- Mathpaid to the people who make it, per hour of work$914 figures
- Lawpaid to the people who make it, per hour of work$905 figures
- Writingpaid to the people who make it, per hour of work$884 figures
- Allpaid to the people who make it, per hour of work$8515 figures
- Biologypaid to the people who make it, per hour of work$848 figures
- Chemistrypaid to the people who make it, per hour of work$827 figures
- Codingpaid to the people who make it, per hour of work$814 figures
- Law medicine financepaid to the people who make it, per hour of work$773 figures
- Red teamingpaid to the people who make it, per hour of work$699 figures
- Languagespaid to the people who make it, per hour of work$403 figures
- Generalistpaid to the people who make it, per hour of work$218 figures
- Codingpaid to the people who make it, per task$2243 figures
- Annotationmade to order, per task$0.06614 figures
- Medicinemade to order, per hour of work$125 to $1532 figures
- Labelingmade to order, per hour of work$73 to $942 figures
- Annotationmade to order, per hour of work$6.00 to $192 figures
- Generalistmade to order, per hour of work$13 to $142 figures
- Preference rankingmade to order, per record$2.584 figures
- Allmade to order, per contract year$150M to $200M2 figures
How the rates are computed
Every figure is one observation: a price somebody published, with its source, its date and its own words, on the price board. Each is converted to a comparable unit by arithmetic on what the source states ($24,500 per dataset of 1,000 hours is $25 per hour of audio), never by estimating a size.
A range counts as its geometric midpoint, because prices in this market span orders of magnitude: $100K to $4M centres near $632K, not $2M. Figures given only as a floor or a ceiling are counted but not averaged. A median is shown from 3 figures; below that, the figures themselves.
The sides are where the money moves: paid to the people who make it; offered to holders; sales that happened; asking prices off the shelf; made to order; cost to produce. They are never mixed in one median, and nor are prices that buy research use only (most academic catalogue fees), which have their own line.
Sources are re-read every day and each figure is checked against its page; a figure that changes is replaced, and the old one is kept as history.
Cite this
AIData.Broker, “What AI training data goes for, on each side of the market”, reviewed 17 Oct 2026. https://www.aidata.markets/rates
Free to reuse under CC BY 4.0 if you credit AIData.Broker and link to this page. Each row names the publication that stated it, and words quoted from a source keep that source's own terms.