Kinds of data

Datasets, tasks and environments

Things built for training and testing models: curated datasets, graded tasks, simulated apps and workplaces.

Public price: $1,200 to $2,500
per environment
Demand: Rising
Who buys: Frontier labs, neolabs, environment marketplaces

The market rate for datasets and environments, by side

How the rates are computed

The median where three or more figures exist, otherwise the figures we hold. Each side is a different point in the market: what the people who make it are paid, what holders are offered, what catalogues ask, what was paid.

Paid to the people who make it

$1,500

per environment, median of 29 figures; middle half $1,200 to $2,500

Paid to the people who make it

$89

per hour of work, median of 12 figures; middle half $64 to $101

Paid to the people who make it

$300

per task, median of 11 figures; middle half $175 to $650

Paid to the people who make it

$200

per demonstration, from 1 figure

Paid to the people who make it

$1,581

per question, from 1 figure

Made to order

$2,837

per task, median of 8 figures; middle half $826 to $5,972

Made to order

$42.7K

per environment, median of 4 figures; middle half $28.3K to $56.9K

Made to order

$187

per hour of work, from 1 figure

Cost to produce

$0.42

per demonstration, median of 6 figures; middle half $0.093 to $0.74

Cost to produce

$1,500

per task, median of 5 figures; middle half $400 to $2,400

What datasets and environments has sold or been offered for

The whole price board

Public prices for datasets and environments
What was sold or offered	Price	When	What to know	Source
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 per task, typically 6 to 7 hours	Oct 2026	Paid only for a task that stumps frontier models; a ceiling, not a typical payout. The same programme's other page (outlier.ai/experts/tbench, held as outlier-tbench-task) still states up to $1,350 per task on the same day, so the two live pages differ by $450; at 6 to 7 hours the ceiling is about $257 to $300 an hour. Undated page, dated by the read.	Outlier AI (with Scale AI), Terminal-Bench expert page The payer's own page
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	Custom Quote ($50K+ band on the quote form) per digital-twin environment (a floor)	Oct 2026	Quoted to order; the quote form's budget band for this tier is the only figure and is a floor. Includes an on-site deployment day.	Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research A vendor's published price guide
Adding a robot model not yet supported (URDF or MJCF and, where needed, dynamics identified from hardware) to a custom robotics RL environment	$2–5K per robot added	Oct 2026	Add-on to the environment tiers; the cost depends on the quality of the robot description file and whether dynamics must be identified from hardware.	Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research A vendor's published price guide
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	From $5,000 (quote form band: $5K to $10K) per environment (1 to 3 tasks)	Oct 2026	The first vendor found that publishes a per-environment price for an RL environment built to order. Delivered as a private repo licensed to the buyer for perpetual internal use; commercial redistribution rights are negotiable. Robotics simulation, not a software or web replica, so compare with care against the $20K website replica.	Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research A vendor's published price guide
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	From $25,000 (quote form band: $25K to $50K) per environment suite (5 to 10 tasks)	Oct 2026	Benchmark-grade suite tier from the same price list as the Starter tier. Five times the Starter floor for about three times the tasks and calibrated dynamics.	Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research A vendor's published price guide
Business operations expert work on Mercor, including briefs and reference answers for simulated work environments (standing talent network band)	$90 to $120 an hour per hour	Oct 2026	The posting closed; its page now shows the standing Mercor talent network for the field, which states the band its listings have paid and names a project's own rate only when it invites someone (was $60 to $100 an hour on the posting).	Mercor, Marketing Expert: Brand Marketing (AI Agent Environments) (closed) and its Business Operations Expert talent network The payer's own page
Compute for one post-training run of an open model (GLM 4.7) on Mercor's 1,800-task APEX Agents dataset	about $500K in compute per post-training run on 1,800 tasks	Oct 2026	Mercor's chief executive on the lift its own task set gave an open model: the corporate-law score went from 4.7% to 26.6% after training on 1,800 tasks. The compute is the lab's training cost, not a price paid for the tasks; the page is undated, so the date is the day it was read.	Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built A vendor's published price guide
Legal-environment RL tasks ordered at scale from Mercor (the most common kind of order)	$2,000 per task per task	Oct 2026	Mercor's chief executive describing a typical order and its scale: some frontier labs might buy 50,000 tasks a month. The page is undated, so the date is the day it was read.	Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built A vendor's published price guide
Cost basis of one expert-written RL task at Mercor (ten hours of expert time at $150 an hour)	a $1,500 cost basis per task	Oct 2026	The vendor's own cost floor for a task before its margin; the margin then depends on how differentiated the task is. The page is undated, so the date is the day it was read.	Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built A vendor's published price guide
RL training tasks sold by Mercor to AI labs, full price range	from $50 to $10,000 per task	Oct 2026	The vendor's own account of what labs pay it per task; the range depends on how differentiated and frontier the task is. The page is undated, so the date is the day it was read.	Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built A vendor's published price guide
Tasks built to a buyer's request on HUD's DataVendor marketplace	around $500 to $1,500 a task per task, with hundreds to thousands of tasks per request	Oct 2026	Buyers post requests (opportunities) and vendors build tasks to them, often from their own codebases or records. Before DataVendor's 20% platform fee at withdrawal. Read from the markdown copy the docs publish, because the rendered page typesets its dollar amounts as math and scrambles them for any reader but a browser; unchanged on 15 Oct 2026.	DataVendor (HUD) docs, What you can sell (the page's markdown) The payer's own page
Recorded AI-coding-assistant session (an agent trajectory of about 50 turns) contributed to a lab via Mercor	$200 per accepted session per accepted session (about 50 turns, 2 to 3 hours)	Oct 2026	Mercor pays builders to do their own project work in a recorded terminal coding assistant for an AI research organisation. Cut from $500 per accepted session (read 11 October 2026) to $200; the page header now reads "$200 / task". Closed: Mercor shows the posting no longer accepting applications (read 2026-10-17).	Mercor, Builders & Developers: Share Your AI Coding Sessions The payer's own page
Software engineer at Mechanize, hired to build environments and evals for frontier coding agents	$300K–$400K salary · equity · bonus per year (base salary, plus equity and bonus)	Oct 2026	Undated live page. Base salary only; the 2025 press figure of $500,000 likely included more than base. Labour cost behind a vendor that sells RL tasks to labs.	Mechanize, Apply page The payer's own page
Terminal-style coding task that stumps frontier models (Outlier, run with Scale AI)	Up to $1,350 USD/task per task (paid to the contributor)	Oct 2026	Contributor payout ceiling, not the lab's price; the page says rates vary by expertise and lower rates apply to non-core work.	Outlier (Scale AI) The payer's own page
Corporate finance professional building RL environments from financial models and narrated walkthroughs (Terac)	Earn $1,000 / task per task (paid to the contributor)	Oct 2026	Contributor pay per task, about 20 hours a week of work; it sits at the middle of the $200 to $2,000 per task band Epoch reports labs pay, so either these tasks are large or the lab price is higher.	Terac, Corporate Finance Professionals: Build RL Enviornments The payer's own page
Machine learning engineer building scenarios and worlds in an RL environment platform (Terac contract)	Earn $90 per hour per hour	Oct 2026	Remote contractor rate paid weekly; the buyer behind Terac is not named.	Terac, Machine Learning Engineers: Scenario Building for Reinforcement Learning The payer's own page
Experienced QuickBooks user building RL environments and writing their tasks for accounting and sales workflows (Terac)	Earn $110 / hr per hour (minimum 35 hours a week)	Oct 2026	A domain user, not an engineer, paid to author environments for a business application; at least two weeks of work.	Terac, QuickBooks Experts: Accounting & Sales Workflows The payer's own page
Evaluator running everyday errands with two AI agents, approving key steps, screen-recording and scoring (micro1, US)	Pay: $25-45/hour per hour (about 1.5 hours per run)	Oct 2026	Computer-use agent trajectories with human approval and scoring; at about 1.5 hours per run that is $37.50 to $67.50 of labour per evaluated run. A twin listing (Everyday AI Agent Tester) carries the same terms.	micro1, AI Agent Evaluator The payer's own page
Agentic genomics task with reference solution for GeneBench Pro, Turing Bioinformatics / Computational Genomics Expert (October posting)	$150 per approved task per approved task	Oct 2026	Five-week contract at 40 hours a week; down from $260 in the 17 Sep 2026 posting of the same project.	AITraining.jobs, Turing Bioinformatics / Computational Genomics Expert (Turing's posting, mirrored) The payer's own page
Reproducible physical-science task an AI agent must complete in a Linux terminal, Turing Physical Sciences Expert	$300 per approved task. Tasks are estimated to take six to eight hours per approved task of six to eight hours	Oct 2026	Turing states both the task price and its own time estimate, so the effective hourly pay for PhD-level work is $37.50 to $50.	AITraining.jobs, Turing Physical Sciences Expert, AI Training (Turing's posting, mirrored) The payer's own page
Scientific-computing task (main problem, subproblems, reference solution, tests) authored from the researcher's own work, AfterQuery Scientific Researcher (Computational Task Author)	$200 per approved task, about 2 hours each per approved task of about 2 hours	Oct 2026	About $100 an hour at AfterQuery's own time estimate, matching the listing's structured $100 hourly pay.	AITraining.jobs, AfterQuery Scientific Researcher, Computational Task Author (AfterQuery's posting, mirrored) The payer's own page
Terminal-based scientific task for Terminal Bench Science (environment, expert solution, automated tests), Turing Scientific Computing Expert	Pay per task - $300 per approved task per approved task	Sep 2026	Five-week contract at 40 hours a week; Outlier pays up to $1,350 for a comparable Terminal-Bench task (already on the board).	AITraining.jobs, Turing Scientific Computing Expert, Physical Sciences (Turing's posting, mirrored) The payer's own page
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 per approved task per approved task	Sep 2026	Minimum of one approved task a day for four weeks, no cap. Turing posts the same project and rate under three other titles (Computer Science, Electrical Engineering, Team Management, 24 Sep to 1 Oct 2026); counted once here.	AITraining.jobs, Turing Software Engineer with Python & Docker experience (Turing's posting, mirrored) The payer's own page
Agentic genomics task with reference solution for GeneBench Pro, Turing Bioinformatics / Computational Genomics Expert (September posting)	$260 per approved task per approved task	Sep 2026	Same project re-posted on 6 Oct 2026 at $150 per approved task (next row): the rate fell 42 percent in three weeks.	AITraining.jobs, Turing Bioinformatics / Computational Genomics Expert (Turing's posting, mirrored) The payer's own page
Research-grade scientific-computing evaluation task authored from the researcher's own computational work, AfterQuery Scientific Computing Research Expert	$300 per approved task, roughly 4 hours of work each per approved task of about 4 hours	Sep 2026	About $75 an hour at AfterQuery's own time estimate, which matches the listing's structured $75 hourly pay; most tasks need a round or two of revision before approval.	AITraining.jobs, AfterQuery Scientific Computing Research Expert (AfterQuery's posting, mirrored) The payer's own page
Auditor of repository-level software-engineering benchmark tasks (reference patches, test harnesses) for Mercor	$70-$90 / hr per hour	Sep 2026	QA labour on verifiable coding tasks, US only; page said posted a month before reading.	Mercor, SWE-Bench Task Auditor The payer's own page
Engineer building the pipeline that captures and scrubs coding-agent (Copilot CLI) traces for a lab dataset (Turing contract)	Compensation: $100/hour per hour (3-month contract, 20+ hours a week)	Sep 2026	Rate quoted from Turing's posting text on a job aggregator (listed 34 days before reading); the aggregator's own header estimate of $25 to $55 is not the stated rate.	aitrainer.work, Turing: Software / Platform / Data Engineer, Agent Trace Collection A catalogue or marketplace listing
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 USD per task per accepted and approved task	Aug 2026	A 3 to 4 week contract paying only on acceptance, one free revision, no cap on volume; the paid pilot task also pays $50. A twin posting limited to India (job 92074380, posted 3 Sep 2026) states the same $50, so it is one programme at one rate. The lowest per-task rate for building an agent environment on record.	Techstars job board, Careerflow: AI/ML Software Engineer, Task Creator (RL Environments) (Contract) The payer's own page
Member of Technical Staff, Deeptune Environments (Mercor), building RL training environments for labs	$160K – $500K • Offers Equity per year (salary)	Aug 2026	Mercor folded Deeptune in as its environments team; the role builds the simulated apps, tool surfaces, tasks and verifiers.	Mercor on Ashby, Member of Technical Staff, Deeptune Environments The payer's own page
One decent RL coding task bought by a frontier lab from a data or environment company	$5000+ per coding task (a floor: 'plus')	Jul 2026	Analyst reporting, not a price list: what labs are willing to pay, stated as a floor. The same piece says Anthropic has been the most aggressive lab in buying coding data from RL environment startups. Ten times Epoch's typical $200 to $2,000 per task and above Mechanize's $500 minimum, so a 2026 coding task of frontier difficulty prices like a small project, not like annotation.	SemiAnalysis, The Future of Meta Superintelligence: A 1 Year Progress Update (Kan, Martin-Prin, Eliahou Ontiveros, Patel) Press report
Labour cost of one good RL coding task at Mechanize: an engineer paid $400k+ a year expected to make one good task a week	$400k+ per year ... 1 good task per week per engineer-year, at one good task per week	Jul 2026	A builder's own production cost per task, reported by SemiAnalysis: salary only, before compute, review and margin. It sits above the $5,000+ labs are said to pay per coding task, so Mechanize's tasks must sell above that floor or be sold more than once.	SemiAnalysis, The Future of Meta Superintelligence: A 1 Year Progress Update Press report
Software Engineer (RL Environments) at AfterQuery	$260K – $290K • Offers Equity • $200K – $400K Bonus per year (base salary; bonus of $200K to $400K on top)	Apr 2026	Base only in low/high; with the stated bonus, cash reaches $460K to $690K a year. The role designs simulations, data and evaluations for frontier labs.	AfterQuery on Ashby, Software Engineer - RL Environments The payer's own page
Reviewing Android agent action sequences and screenshots, pass or fail, on the contributor's own phone (micro1)	pays $0.216 per task (about 1 minute each)	Feb 2026	Piece rate stated in the posting's screening question; 60 tasks earn about $13, so roughly $13 an hour. Listing closed by the day read.	micro1, Mobile Agent QA Evaluator (Android) The payer's own page
Human test-taker calibrating ARC-AGI-2 tasks in a 90-minute session	$115-150 for participation in a 90-minute test session, plus a $5 incentive reward per correctly completed task per 90-minute session (plus $5 per correct task)	Jan 2026	Generalist human calibration, not expert authoring; the per-task bonus rewards solved items.	arXiv, ARC-AGI-2 (v2) The payer's own page
RL environment supply contract between a vendor and a smaller (neo) lab	contracts in the $300k-$500k range per contract (often quarterly)	Jan 2026	One researcher's observation; Epoch says contract sizes are often six to seven figures per quarter and vary with the number of tasks.	Epoch AI Press report
One especially complex software-engineering RL task with its grader (upper end reported to Epoch AI)	$20k per task would be rare but possible per task	Jan 2026	The ceiling interviewees gave for a single task; the same piece puts the usual range at $200 to $2,000 per task. Rare, and only for complex software-engineering tasks.	Epoch AI, An FAQ on reinforcement learning environments (Denain and Barber) Press report
Exclusive rights instead of shared rights	Roughly 4 to 5 times the non-exclusive price multiplier	Jan 2026	Two founders said this independently, about environments. It is the largest single lever on price, and the one a single-buyer offer never shows you.	Epoch AI Press report
One training task with its grader	$200 to $2,000 per task	Jan 2026	From interviews with 18 people who build and buy them. $20,000 for a single task is rare but happens.	Epoch AI Press report
A working replica of a website or app for an agent to practise in	About $20K; about $300K for a Slack-grade clone per environment	Jan 2026	Where a replica can be built, nobody pays the original site's owner. That is the ceiling on what interface data is worth.	Epoch AI Press report
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')	EUR 968,521 (lot estimate EUR 1,080,000) per contract (24 months)	Dec 2025	A public buyer's award for agent-evaluation infrastructure made to order, signed 2025-11-27 and won about 10% under the EUR 1,080,000 estimate (contract notice 272332-2025) by a French public consortium including LNE. It buys the harness to run agentic environments, not a set of tasks, so compare it with environment-platform builds rather than per-task prices. The lot's scope is quoted from the contract notice (https://ted.europa.eu/en/notice/272332-2025/xml): 'interaction with digital environments such as a browser, command line, or a full operating system'. Sister lots in the same award: cyber offence EUR 1,700,000 (Capgemini, Wavestone, LNE) and loss of control EUR 1,167,484 (EquiStamp, METR, Epoch AI), both mainly risk modelling.	TED (EU Official Journal), contract award notice 864574-2025: Artificial Intelligence Act, technical assistance for AI safety (European Commission, DG CNECT) The payer's own page
Software engineering intern at Mechanize designing RL tasks (prompt, environment and automated grader)	$150 per hour per hour	Nov 2025	Intern rate, already above most expert contractor rates for environment work; each task is a self-contained software challenge with its own grader.	Mechanize on Ashby, Software Engineering Intern The payer's own page
Open-source RL environment built for Prime Intellect's Environments Hub (bounty)	$100-500 for open-access bounties; about $1,000 to $5,000+ for application-only bounties per environment	Oct 2025	Bounties for open, non-exclusive environments released publicly, so far below private exclusive contracts; the larger tier's figures render garbled on the page.	Prime Intellect The payer's own page
Administration of the UK AI Safety Institute's fund for sourcing external evaluations (programs that test frontier models' dangerous capabilities), awarded to Smarter Society	£2,300,000 awarded (notice value up to £5,300,000) per contract (Oct 2024 to Mar 2025)	Oct 2025	The vehicle behind AISI's evals bounty (GBP 3,000 to 15,000 per successful task, store id aisi-evals-bounty-per-successful-task). The notice does not split the administrator's fee from the money paid out to evaluation builders, so read it as the size of the sourcing programme, not a price per task.	UK Contracts Finder (official API record), DSIT award notice: AISI Evaluation Sourcing Fund Administration The payer's own page
Human computer-use demonstrations across 3 operating systems and 200+ apps (OpenCUA AgentNet, 22K tasks)	about USD 20,000 per dataset of 22K annotated tasks	Oct 2025	Part-time students, Prolific and four annotation companies over six months at about ten tasks an hour; total dataset cost was about $32,000 with reasoning synthesis.	arXiv, OpenCUA (v3), Appendix C.2 The payer's own page
Synthetic chain-of-thought reasoning added to each human computer-use demonstration (OpenCUA)	USD 0.6 per task per task	Oct 2025	Model cost of enriching a human trajectory; it is two thirds of what the human demonstration itself cost.	arXiv, OpenCUA (v3), Appendix C.2 The payer's own page
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)	£545,000 per contract (Dec 2024 to Sep 2025)	Oct 2025	Private (non-public) CTF environments bought by an evaluator over ten months; the count of CTFs is not stated. The supplier is written as 'Rethinks Priorities' in the notice.	UK Contracts Finder (official API record), DSIT award notice: prj_3902, WP 2, Private Capture the Flag Evaluations The payer's own page
AndroidWorld mobile-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)	Android World: $1,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
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 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
BFCL-v3 (Berkeley Function Calling Leaderboard) reimplemented as an environment: tool and function calling (Prime Intellect Environments Hub, application-only bounty)	$500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Biomni-R0 (Stanford) reimplemented as an environment: a biomedical research agent (Prime Intellect Environments Hub, application-only bounty)	$1,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Box (Prime Intellect Environments Hub, application-only bounty)	Box: $2,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Environment synthesised and validated from a textbook: CLRS, Introduction to Algorithms (Prime Intellect Environments Hub, application-only bounty)	$1,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
ComputerRL (Zhipu) computer-use RL setup reimplemented as an environment (Prime Intellect Environments Hub, application-only bounty)	ComputerRL: $5,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Software-library environment for agent tasks in Django (Prime Intellect Environments Hub, application-only bounty)	Django: $3,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Figma (Prime Intellect Environments Hub, application-only bounty)	Figma: $2,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Intercom (Prime Intellect Environments Hub, application-only bounty)	Intercom: $1,200 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Mind2Web web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)	Mind2Web: $2,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MLE-Bench (OpenAI) reimplemented as an environment: ML engineering competitions for an agent (Prime Intellect Environments Hub, application-only bounty)	$1,200 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Monday.com (Prime Intellect Environments Hub, application-only bounty)	Monday: $1,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
nanoGPT speedrunning environment: an agent trains a small language model as fast as it can (Prime Intellect Environments Hub, application-only bounty)	$2,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Vercel's Next.js eval suite implemented as an environment (Prime Intellect Environments Hub, application-only bounty)	Next.js Bench: $3,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Notion (Prime Intellect Environments Hub, application-only bounty)	Notion: $1,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
OpenHands agent framework integration with the Environments Hub (Prime Intellect Environments Hub, application-only bounty)	$2,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
OSWorld computer-use benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)	OSWorld: $1,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
PaperBench (OpenAI) reimplemented as an environment: an agent replicates research papers (Prime Intellect Environments Hub, application-only bounty)	$1,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Software-library environment for agent tasks in PyTorch (Prime Intellect Environments Hub, application-only bounty)	PyTorch: $5,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Shop-R1 reimplemented as an environment: an agent simulating online shopping behaviour (Prime Intellect Environments Hub, application-only bounty)	$2,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Stripe (Prime Intellect Environments Hub, application-only bounty)	Stripe: $1,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
SWE-bench Verified coding benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)	SWE-bench Verified: $3,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
TAU2 (Sierra) extended training set generation pipeline: customer-service tool-use tasks generated for training (Prime Intellect Environments Hub, application-only bounty)	$1,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
WebArena web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)	WebArena: $2,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Webflow (Prime Intellect Environments Hub, application-only bounty)	WebFlow: $2,500 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
WebVoyager web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)	WebVoyager: $1,200 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
MCP backend simulator and agent environment for Zapier (Prime Intellect Environments Hub, application-only bounty)	Zapier: $2,000 per environment (bounty)	Oct 2025	Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.	Prime Intellect Environments Program bounty sheet, Application-Only tab The payer's own page
Anthropic's discussed spend on RL environments bought from vendors over the following year	more than $1 billion total over the next year	Sep 2025	TechCrunch citing The Information; a discussed budget, not a signed contract. Read with the $200 to $2,000 per task band: at those prices it is roughly 0.5 to 5 million tasks.	TechCrunch, Silicon Valley bets big on environments to train AI agents Press report
Software engineer salary at Mechanize for building RL environments (2025 report)	$500,000 salaries per year (salary)	Sep 2025	Employee salary, not a contractor rate; TechCrunch contrasts it with hourly contractor pay at Scale AI or Surge.	TechCrunch, Silicon Valley bets big on environments to train AI agents Press report
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 per RL task	Aug 2025	An argument from compute cost by a company that sells RL environments, not an observed price: about $480 of compute per task per training run, and $2,400 over a task's life at five reuses.	Mechanize A vendor's published price guide
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 for each RL task per RL task, compute over its life	Aug 2025	A modelled compute cost a lab bears when it trains on one task, not a price paid for the task. Written by a company that sells RL environments, to argue labs should spend at least about $500 per task.	Mechanize, Cheap RL tasks will waste compute A vendor's published price guide
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)	£999,600 per contract (Jan to Apr 2025)	Jun 2025	A government buyer's published award for an environment made to its order: the requirement names one complex cyber range environment plus its tasks, solvers, graders and human baselines, so this is the nearest public figure to a per-environment price for a high-fidelity private environment. Awarded 2025-01-10 by call-off from a dynamic purchasing system, contract 2025-03-13 to 2025-04-30.	UK Contracts Finder (official API record), DSIT award notice: WP 1: Cyber Range Scenarios, Enterprise Networks The payer's own page
Novel research-level maths problem with solution written for the FrontierMath benchmark	up to $1,000 apiece per problem	Jun 2025	Paid by Epoch AI to professional mathematicians; problems were graded into three difficulty tiers.	Financial Times (Anjana Ahuja, republished by LIMS) Press report
Synthetic multimodal web-agent trajectory from exploration (Explorer, 94K successful trajectories)	28 cents per successful trajectory per successful trajectory	May 2025	Producer's own compute and API cost, no humans; over 94K successful trajectories were produced.	arXiv, Explorer (v4) The payer's own page
Human expert baseline attempt on an 8-hour AI R&D task (METR RE-Bench)	approximately $1,855 per expert attempt (8-hour task)	May 2025	Average paid per expert run; over the 8-hour budget this is about $232 an hour. An agent run of the same length cost about $123 in tokens.	arXiv, RE-Bench (v2) The payer's own page
150,000 synthetic web-agent trajectories with judge annotations, generated on 1,200 V100 GPU hours (InSTA)	costs $521.55 per dataset of 150k trajectories	May 2025	Open models (Qwen3) on spot GPUs, no human annotation; the authors call it a fraction of what industry labs spend on agents.	arXiv, InSTA (v2) The payer's own page
One 12-hour agent rollout replicating a research paper (o1 IterativeAgent on PaperBench)	$400 in API credits per rollout on one paper	Apr 2025	Inference cost of one long-horizon trajectory, plus about $66 to grade it; it shows why labs pay for tasks where a long rollout teaches something.	arXiv, PaperBench (v3) The payer's own page
Human baseliner timing software and research tasks for METR's HCAST benchmark	between $50 - $100 per hour, plus $25 - $150 per hour in performance bonuses per hour (base, before bonuses)	Mar 2025	With bonuses for fast, successful completion the total runs $75 to $250 an hour.	arXiv, HCAST The payer's own page
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)	£590,500 per contract (Jan to Jun 2025)	Mar 2025	Sister work package to the SpecOps enterprise-network range: the same buyer paid a second vendor for critical-infrastructure range scenarios. The notice does not state how many scenarios, so no per-environment figure. Awarded 2024-12-12, contract 2025-01-02 to 2025-06-30.	UK Contracts Finder (official API record), DSIT award notice: Cyber Evaluations, WP1, Cyber Range Scenarios The payer's own page
Synthetic GUI/web agent trajectory generated by replaying web tutorials (AgentTrek)	$0.551 per trajectory per trajectory	Mar 2025	Model API cost with no human annotators; the abstract rounds it to $0.55. Executing 1,000 tasks with GPT-4o cost about $215 in the same paper.	arXiv, AgentTrek (v2) The payer's own page
Real freelance software tasks from Upwork used as the SWE-Lancer eval set	$50 bug fixes to $32,000 feature implementations; $1 million total across 1,400+ tasks per task (original freelance payout)	Feb 2025	These are what clients originally paid freelancers for each task, used by OpenAI to value the eval; roughly $700 per task on average (derived).	OpenAI (SWE-Lancer paper, arXiv) The payer's own page
Human-annotated web-agent trajectory in the Mind2Web training set (as costed by the Synatra authors)	each trajectory costing $0.8 per trajectory	Nov 2024	A third party's costing of an academic crowd-annotated dataset; the per-trajectory figure is the academic floor for human computer-use demonstrations.	arXiv, Synatra (v2), comparison with Mind2Web The payer's own page
Synthetic web-agent demonstration converted from tutorials and indirect knowledge (Synatra)	$0.031 each per demonstration	Nov 2024	The authors put it at 3% of the cost of a human demonstration and found it can beat an equal number of human demonstrations from limited domains.	arXiv, Synatra (v2) The payer's own page
Development time on an evaluation task for the UK AI Safety Institute (evals bounty indication)	£100-200 per development hour per development hour	Nov 2024	The rate the buyer uses to size a task bounty, paid only for tasks that succeed; stage-two applicants also get a GBP 2,000 award toward compute.	UK AI Security Institute, evals bounty (blog) The payer's own page
One successful dangerous-capability or autonomous-agent evaluation task built for the UK AI Safety Institute (evals bounty)	£3000-£15,000 for a successful task per successful task	Nov 2024	A government evaluator buying tasks directly from builders (individuals or organisations), paid in full only once the evaluation meets its criteria; exceptional work may earn more.	UK AI Security Institute, evals bounty (blog) The payer's own page
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 per hour of professional time the task takes to complete	Nov 2024	The later rule of METR's task bounty (paused to new submissions on 2024-11-18, over $100,000 paid out): a 10-hour task would earn $3,000. Per-task value depends on task length, which varies, so no per-task figure is derived.	METR Task Development Guide, Introduction The payer's own page
Expert-written benchmark question for Humanity's Last Exam (eval set)	$5,000 each for the top 50 questions, $500 each for the next 500 per accepted question	Sep 2024	Prize money from a $500,000 pool paid by the eval's builders (CAIS and Scale AI) for questions frontier models could not answer.	Center for AI Safety The payer's own page
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)	£440,000 per contract (Jun to Oct 2024)	Aug 2024	Four-month build of private agent scenarios; scenario count not stated. The supplier is named 'Grey Swan Security Inc'; the agent-safeguard benchmark work matches Gray Swan AI's line of business, but the notice alone does not confirm it is the same company.	UK Contracts Finder (official API record), DSIT award notice: offensive cyber capability and agent safeguards measurement The payer's own page
Suites of tasks ('evals') for evaluating AI systems, created to order (Rethink Priorities for the UK AI Safety Institute)	£232,517.64 (first awarded at £155,011.76) per contract (Jan to Mar 2024)	Feb 2024	Twelve-week task-suite build, extended in time and value by half again (a sign that a first task order grows once the buyer sees the tasks). Task count not stated.	UK Contracts Finder (official API record), DSIT award notice: Creating suites of tasks ('evals') to use in the evaluation of AI systems The payer's own page
One private agent-evaluation task implementation for METR that takes an experienced engineer about 10 hours (worked example)	the base price would be $3000 ... pay $4500 overall per task implementation	Dec 2023	METR, an evaluator that uses the tasks itself, pays the builder directly with no vendor in between: a base of 3x the human cost to solve the task, plus up to a 50% bonus. Tasks must stay off the public web and out of training data.	METR, Bounty: Diverse hard tasks for LLM agents The payer's own page
Crowd annotator demonstrating web tasks for the Mind2Web agent dataset	estimated hourly rate of $10.10 per hour	Dec 2023	Academic rate set to Ohio minimum wage; the human end of the trajectory market before labs paid experts.	arXiv, Mind2Web (v3) The payer's own page
PhD-level expert writing and validating GPQA benchmark questions	approximately $95 per hour per hour (average; maximum possible $150)	Nov 2023	Paid as a $10 base per question plus bonuses when experts agree and non-experts fail, so pay tracks question difficulty.	arXiv, GPQA The payer's own page

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
Price
Up to $1,800 USD/task
per task, typically 6 to 7 hours
When
Oct 2026
What to know
Paid only for a task that stumps frontier models; a ceiling, not a typical payout. The same programme's other page (outlier.ai/experts/tbench, held as outlier-tbench-task) still states up to $1,350 per task on the same day, so the two live pages differ by $450; at 6 to 7 hours the ceiling is about $257 to $300 an hour. Undated page, dated by the read.
Source
Outlier AI (with Scale AI), Terminal-Bench expert page
The payer's own page
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
Price
Custom Quote ($50K+ band on the quote form)
per digital-twin environment (a floor)
When
Oct 2026
What to know
Quoted to order; the quote form's budget band for this tier is the only figure and is a floor. Includes an on-site deployment day.
Source
Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research
A vendor's published price guide
Adding a robot model not yet supported (URDF or MJCF and, where needed, dynamics identified from hardware) to a custom robotics RL environment
Price
$2–5K
per robot added
When
Oct 2026
What to know
Add-on to the environment tiers; the cost depends on the quality of the robot description file and whether dynamics must be identified from hardware.
Source
Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research
A vendor's published price guide
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
Price
From $5,000 (quote form band: $5K to $10K)
per environment (1 to 3 tasks)
When
Oct 2026
What to know
The first vendor found that publishes a per-environment price for an RL environment built to order. Delivered as a private repo licensed to the buyer for perpetual internal use; commercial redistribution rights are negotiable. Robotics simulation, not a software or web replica, so compare with care against the $20K website replica.
Source
Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research
A vendor's published price guide
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
Price
From $25,000 (quote form band: $25K to $50K)
per environment suite (5 to 10 tasks)
When
Oct 2026
What to know
Benchmark-grade suite tier from the same price list as the Starter tier. Five times the Starter floor for about three times the tasks and calibrated dynamics.
Source
Robotics Center (RCSV), Custom Reinforcement Learning Environments for Robotics Research
A vendor's published price guide
Business operations expert work on Mercor, including briefs and reference answers for simulated work environments (standing talent network band)
Price
$90 to $120 an hour
per hour
When
Oct 2026
What to know
The posting closed; its page now shows the standing Mercor talent network for the field, which states the band its listings have paid and names a project's own rate only when it invites someone (was $60 to $100 an hour on the posting).
Source
Mercor, Marketing Expert: Brand Marketing (AI Agent Environments) (closed) and its Business Operations Expert talent network
The payer's own page
Compute for one post-training run of an open model (GLM 4.7) on Mercor's 1,800-task APEX Agents dataset
Price
about $500K in compute
per post-training run on 1,800 tasks
When
Oct 2026
What to know
Mercor's chief executive on the lift its own task set gave an open model: the corporate-law score went from 4.7% to 26.6% after training on 1,800 tasks. The compute is the lab's training cost, not a price paid for the tasks; the page is undated, so the date is the day it was read.
Source
Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built
A vendor's published price guide
Legal-environment RL tasks ordered at scale from Mercor (the most common kind of order)
Price
$2,000 per task
per task
When
Oct 2026
What to know
Mercor's chief executive describing a typical order and its scale: some frontier labs might buy 50,000 tasks a month. The page is undated, so the date is the day it was read.
Source
Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built
A vendor's published price guide
Cost basis of one expert-written RL task at Mercor (ten hours of expert time at $150 an hour)
Price
a $1,500 cost basis
per task
When
Oct 2026
What to know
The vendor's own cost floor for a task before its margin; the margin then depends on how differentiated the task is. The page is undated, so the date is the day it was read.
Source
Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built
A vendor's published price guide
RL training tasks sold by Mercor to AI labs, full price range
Price
from $50 to $10,000
per task
When
Oct 2026
What to know
The vendor's own account of what labs pay it per task; the range depends on how differentiated and frontier the task is. The page is undated, so the date is the day it was read.
Source
Sequoia Capital podcast, Brendan Foody (Mercor): How RL environments are built
A vendor's published price guide
Tasks built to a buyer's request on HUD's DataVendor marketplace
Price
around $500 to $1,500 a task
per task, with hundreds to thousands of tasks per request
When
Oct 2026
What to know
Buyers post requests (opportunities) and vendors build tasks to them, often from their own codebases or records. Before DataVendor's 20% platform fee at withdrawal. Read from the markdown copy the docs publish, because the rendered page typesets its dollar amounts as math and scrambles them for any reader but a browser; unchanged on 15 Oct 2026.
Source
DataVendor (HUD) docs, What you can sell (the page's markdown)
The payer's own page
Recorded AI-coding-assistant session (an agent trajectory of about 50 turns) contributed to a lab via Mercor
Price
$200 per accepted session
per accepted session (about 50 turns, 2 to 3 hours)
When
Oct 2026
What to know
Mercor pays builders to do their own project work in a recorded terminal coding assistant for an AI research organisation. Cut from $500 per accepted session (read 11 October 2026) to $200; the page header now reads "$200 / task". Closed: Mercor shows the posting no longer accepting applications (read 2026-10-17).
Source
Mercor, Builders & Developers: Share Your AI Coding Sessions
The payer's own page
Software engineer at Mechanize, hired to build environments and evals for frontier coding agents
Price
$300K–$400K salary · equity · bonus
per year (base salary, plus equity and bonus)
When
Oct 2026
What to know
Undated live page. Base salary only; the 2025 press figure of $500,000 likely included more than base. Labour cost behind a vendor that sells RL tasks to labs.
Source
Mechanize, Apply page
The payer's own page
Terminal-style coding task that stumps frontier models (Outlier, run with Scale AI)
Price
Up to $1,350 USD/task
per task (paid to the contributor)
When
Oct 2026
What to know
Contributor payout ceiling, not the lab's price; the page says rates vary by expertise and lower rates apply to non-core work.
Source
Outlier (Scale AI)
The payer's own page
Corporate finance professional building RL environments from financial models and narrated walkthroughs (Terac)
Price
Earn $1,000 / task
per task (paid to the contributor)
When
Oct 2026
What to know
Contributor pay per task, about 20 hours a week of work; it sits at the middle of the $200 to $2,000 per task band Epoch reports labs pay, so either these tasks are large or the lab price is higher.
Source
Terac, Corporate Finance Professionals: Build RL Enviornments
The payer's own page
Machine learning engineer building scenarios and worlds in an RL environment platform (Terac contract)
Price
Earn $90 per hour
per hour
When
Oct 2026
What to know
Remote contractor rate paid weekly; the buyer behind Terac is not named.
Source
Terac, Machine Learning Engineers: Scenario Building for Reinforcement Learning
The payer's own page
Experienced QuickBooks user building RL environments and writing their tasks for accounting and sales workflows (Terac)
Price
Earn $110 / hr
per hour (minimum 35 hours a week)
When
Oct 2026
What to know
A domain user, not an engineer, paid to author environments for a business application; at least two weeks of work.
Source
Terac, QuickBooks Experts: Accounting & Sales Workflows
The payer's own page
Evaluator running everyday errands with two AI agents, approving key steps, screen-recording and scoring (micro1, US)
Price
Pay: $25-45/hour
per hour (about 1.5 hours per run)
When
Oct 2026
What to know
Computer-use agent trajectories with human approval and scoring; at about 1.5 hours per run that is $37.50 to $67.50 of labour per evaluated run. A twin listing (Everyday AI Agent Tester) carries the same terms.
Source
micro1, AI Agent Evaluator
The payer's own page
Agentic genomics task with reference solution for GeneBench Pro, Turing Bioinformatics / Computational Genomics Expert (October posting)
Price
$150 per approved task
per approved task
When
Oct 2026
What to know
Five-week contract at 40 hours a week; down from $260 in the 17 Sep 2026 posting of the same project.
Source
AITraining.jobs, Turing Bioinformatics / Computational Genomics Expert (Turing's posting, mirrored)
The payer's own page
Reproducible physical-science task an AI agent must complete in a Linux terminal, Turing Physical Sciences Expert
Price
$300 per approved task. Tasks are estimated to take six to eight hours
per approved task of six to eight hours
When
Oct 2026
What to know
Turing states both the task price and its own time estimate, so the effective hourly pay for PhD-level work is $37.50 to $50.
Source
AITraining.jobs, Turing Physical Sciences Expert, AI Training (Turing's posting, mirrored)
The payer's own page
Scientific-computing task (main problem, subproblems, reference solution, tests) authored from the researcher's own work, AfterQuery Scientific Researcher (Computational Task Author)
Price
$200 per approved task, about 2 hours each
per approved task of about 2 hours
When
Oct 2026
What to know
About $100 an hour at AfterQuery's own time estimate, matching the listing's structured $100 hourly pay.
Source
AITraining.jobs, AfterQuery Scientific Researcher, Computational Task Author (AfterQuery's posting, mirrored)
The payer's own page
Terminal-based scientific task for Terminal Bench Science (environment, expert solution, automated tests), Turing Scientific Computing Expert
Price
Pay per task - $300 per approved task
per approved task
When
Sep 2026
What to know
Five-week contract at 40 hours a week; Outlier pays up to $1,350 for a comparable Terminal-Bench task (already on the board).
Source
AITraining.jobs, Turing Scientific Computing Expert, Physical Sciences (Turing's posting, mirrored)
The payer's own page
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
Price
$300 per approved task
per approved task
When
Sep 2026
What to know
Minimum of one approved task a day for four weeks, no cap. Turing posts the same project and rate under three other titles (Computer Science, Electrical Engineering, Team Management, 24 Sep to 1 Oct 2026); counted once here.
Source
AITraining.jobs, Turing Software Engineer with Python & Docker experience (Turing's posting, mirrored)
The payer's own page
Agentic genomics task with reference solution for GeneBench Pro, Turing Bioinformatics / Computational Genomics Expert (September posting)
Price
$260 per approved task
per approved task
When
Sep 2026
What to know
Same project re-posted on 6 Oct 2026 at $150 per approved task (next row): the rate fell 42 percent in three weeks.
Source
AITraining.jobs, Turing Bioinformatics / Computational Genomics Expert (Turing's posting, mirrored)
The payer's own page
Research-grade scientific-computing evaluation task authored from the researcher's own computational work, AfterQuery Scientific Computing Research Expert
Price
$300 per approved task, roughly 4 hours of work each
per approved task of about 4 hours
When
Sep 2026
What to know
About $75 an hour at AfterQuery's own time estimate, which matches the listing's structured $75 hourly pay; most tasks need a round or two of revision before approval.
Source
AITraining.jobs, AfterQuery Scientific Computing Research Expert (AfterQuery's posting, mirrored)
The payer's own page
Auditor of repository-level software-engineering benchmark tasks (reference patches, test harnesses) for Mercor
Price
$70-$90 / hr
per hour
When
Sep 2026
What to know
QA labour on verifiable coding tasks, US only; page said posted a month before reading.
Source
Mercor, SWE-Bench Task Auditor
The payer's own page
Engineer building the pipeline that captures and scrubs coding-agent (Copilot CLI) traces for a lab dataset (Turing contract)
Price
Compensation: $100/hour
per hour (3-month contract, 20+ hours a week)
When
Sep 2026
What to know
Rate quoted from Turing's posting text on a job aggregator (listed 34 days before reading); the aggregator's own header estimate of $25 to $55 is not the stated rate.
Source
aitrainer.work, Turing: Software / Platform / Data Engineer, Agent Trace Collection
A catalogue or marketplace listing
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)
Price
$50 USD per task
per accepted and approved task
When
Aug 2026
What to know
A 3 to 4 week contract paying only on acceptance, one free revision, no cap on volume; the paid pilot task also pays $50. A twin posting limited to India (job 92074380, posted 3 Sep 2026) states the same $50, so it is one programme at one rate. The lowest per-task rate for building an agent environment on record.
Source
Techstars job board, Careerflow: AI/ML Software Engineer, Task Creator (RL Environments) (Contract)
The payer's own page
Member of Technical Staff, Deeptune Environments (Mercor), building RL training environments for labs
Price
$160K – $500K • Offers Equity
per year (salary)
When
Aug 2026
What to know
Mercor folded Deeptune in as its environments team; the role builds the simulated apps, tool surfaces, tasks and verifiers.
Source
Mercor on Ashby, Member of Technical Staff, Deeptune Environments
The payer's own page
One decent RL coding task bought by a frontier lab from a data or environment company
Price
$5000+
per coding task (a floor: 'plus')
When
Jul 2026
What to know
Analyst reporting, not a price list: what labs are willing to pay, stated as a floor. The same piece says Anthropic has been the most aggressive lab in buying coding data from RL environment startups. Ten times Epoch's typical $200 to $2,000 per task and above Mechanize's $500 minimum, so a 2026 coding task of frontier difficulty prices like a small project, not like annotation.
Source
SemiAnalysis, The Future of Meta Superintelligence: A 1 Year Progress Update (Kan, Martin-Prin, Eliahou Ontiveros, Patel)
Press report
Labour cost of one good RL coding task at Mechanize: an engineer paid $400k+ a year expected to make one good task a week
Price
$400k+ per year ... 1 good task per week
per engineer-year, at one good task per week
When
Jul 2026
What to know
A builder's own production cost per task, reported by SemiAnalysis: salary only, before compute, review and margin. It sits above the $5,000+ labs are said to pay per coding task, so Mechanize's tasks must sell above that floor or be sold more than once.
Source
SemiAnalysis, The Future of Meta Superintelligence: A 1 Year Progress Update
Press report
Software Engineer (RL Environments) at AfterQuery
Price
$260K – $290K • Offers Equity • $200K – $400K Bonus
per year (base salary; bonus of $200K to $400K on top)
When
Apr 2026
What to know
Base only in low/high; with the stated bonus, cash reaches $460K to $690K a year. The role designs simulations, data and evaluations for frontier labs.
Source
AfterQuery on Ashby, Software Engineer - RL Environments
The payer's own page
Reviewing Android agent action sequences and screenshots, pass or fail, on the contributor's own phone (micro1)
Price
pays $0.216
per task (about 1 minute each)
When
Feb 2026
What to know
Piece rate stated in the posting's screening question; 60 tasks earn about $13, so roughly $13 an hour. Listing closed by the day read.
Source
micro1, Mobile Agent QA Evaluator (Android)
The payer's own page
Human test-taker calibrating ARC-AGI-2 tasks in a 90-minute session
Price
$115-150 for participation in a 90-minute test session, plus a $5 incentive reward per correctly completed task
per 90-minute session (plus $5 per correct task)
When
Jan 2026
What to know
Generalist human calibration, not expert authoring; the per-task bonus rewards solved items.
Source
arXiv, ARC-AGI-2 (v2)
The payer's own page
RL environment supply contract between a vendor and a smaller (neo) lab
Price
contracts in the $300k-$500k range
per contract (often quarterly)
When
Jan 2026
What to know
One researcher's observation; Epoch says contract sizes are often six to seven figures per quarter and vary with the number of tasks.
Source
Epoch AI
Press report
One especially complex software-engineering RL task with its grader (upper end reported to Epoch AI)
Price
$20k per task would be rare but possible
per task
When
Jan 2026
What to know
The ceiling interviewees gave for a single task; the same piece puts the usual range at $200 to $2,000 per task. Rare, and only for complex software-engineering tasks.
Source
Epoch AI, An FAQ on reinforcement learning environments (Denain and Barber)
Press report
Exclusive rights instead of shared rights
Price
Roughly 4 to 5 times the non-exclusive price
multiplier
When
Jan 2026
What to know
Two founders said this independently, about environments. It is the largest single lever on price, and the one a single-buyer offer never shows you.
Source
Epoch AI
Press report
One training task with its grader
Price
$200 to $2,000
per task
When
Jan 2026
What to know
From interviews with 18 people who build and buy them. $20,000 for a single task is rare but happens.
Source
Epoch AI
Press report
A working replica of a website or app for an agent to practise in
Price
About $20K; about $300K for a Slack-grade clone
per environment
When
Jan 2026
What to know
Where a replica can be built, nobody pays the original site's owner. That is the ceiling on what interface data is worth.
Source
Epoch AI
Press report
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')
Price
EUR 968,521 (lot estimate EUR 1,080,000)
per contract (24 months)
When
Dec 2025
What to know
A public buyer's award for agent-evaluation infrastructure made to order, signed 2025-11-27 and won about 10% under the EUR 1,080,000 estimate (contract notice 272332-2025) by a French public consortium including LNE. It buys the harness to run agentic environments, not a set of tasks, so compare it with environment-platform builds rather than per-task prices. The lot's scope is quoted from the contract notice (https://ted.europa.eu/en/notice/272332-2025/xml): 'interaction with digital environments such as a browser, command line, or a full operating system'. Sister lots in the same award: cyber offence EUR 1,700,000 (Capgemini, Wavestone, LNE) and loss of control EUR 1,167,484 (EquiStamp, METR, Epoch AI), both mainly risk modelling.
Source
TED (EU Official Journal), contract award notice 864574-2025: Artificial Intelligence Act, technical assistance for AI safety (European Commission, DG CNECT)
The payer's own page
Software engineering intern at Mechanize designing RL tasks (prompt, environment and automated grader)
Price
$150 per hour
per hour
When
Nov 2025
What to know
Intern rate, already above most expert contractor rates for environment work; each task is a self-contained software challenge with its own grader.
Source
Mechanize on Ashby, Software Engineering Intern
The payer's own page
Open-source RL environment built for Prime Intellect's Environments Hub (bounty)
Price
$100-500 for open-access bounties; about $1,000 to $5,000+ for application-only bounties
per environment
When
Oct 2025
What to know
Bounties for open, non-exclusive environments released publicly, so far below private exclusive contracts; the larger tier's figures render garbled on the page.
Source
Prime Intellect
The payer's own page
Administration of the UK AI Safety Institute's fund for sourcing external evaluations (programs that test frontier models' dangerous capabilities), awarded to Smarter Society
Price
£2,300,000 awarded (notice value up to £5,300,000)
per contract (Oct 2024 to Mar 2025)
When
Oct 2025
What to know
The vehicle behind AISI's evals bounty (GBP 3,000 to 15,000 per successful task, store id aisi-evals-bounty-per-successful-task). The notice does not split the administrator's fee from the money paid out to evaluation builders, so read it as the size of the sourcing programme, not a price per task.
Source
UK Contracts Finder (official API record), DSIT award notice: AISI Evaluation Sourcing Fund Administration
The payer's own page
Human computer-use demonstrations across 3 operating systems and 200+ apps (OpenCUA AgentNet, 22K tasks)
Price
about USD 20,000
per dataset of 22K annotated tasks
When
Oct 2025
What to know
Part-time students, Prolific and four annotation companies over six months at about ten tasks an hour; total dataset cost was about $32,000 with reasoning synthesis.
Source
arXiv, OpenCUA (v3), Appendix C.2
The payer's own page
Synthetic chain-of-thought reasoning added to each human computer-use demonstration (OpenCUA)
Price
USD 0.6 per task
per task
When
Oct 2025
What to know
Model cost of enriching a human trajectory; it is two thirds of what the human demonstration itself cost.
Source
arXiv, OpenCUA (v3), Appendix C.2
The payer's own page
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)
Price
£545,000
per contract (Dec 2024 to Sep 2025)
When
Oct 2025
What to know
Private (non-public) CTF environments bought by an evaluator over ten months; the count of CTFs is not stated. The supplier is written as 'Rethinks Priorities' in the notice.
Source
UK Contracts Finder (official API record), DSIT award notice: prj_3902, WP 2, Private Capture the Flag Evaluations
The payer's own page
AndroidWorld mobile-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
Android World: $1,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
AppWorld (Ai2) reimplemented as an environment: an agent operating a world of simulated everyday apps through APIs (Prime Intellect Environments Hub, application-only bounty)
Price
$1,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
BFCL-v3 (Berkeley Function Calling Leaderboard) reimplemented as an environment: tool and function calling (Prime Intellect Environments Hub, application-only bounty)
Price
$500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Biomni-R0 (Stanford) reimplemented as an environment: a biomedical research agent (Prime Intellect Environments Hub, application-only bounty)
Price
$1,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Box (Prime Intellect Environments Hub, application-only bounty)
Price
Box: $2,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Environment synthesised and validated from a textbook: CLRS, Introduction to Algorithms (Prime Intellect Environments Hub, application-only bounty)
Price
$1,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
ComputerRL (Zhipu) computer-use RL setup reimplemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
ComputerRL: $5,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Software-library environment for agent tasks in Django (Prime Intellect Environments Hub, application-only bounty)
Price
Django: $3,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Figma (Prime Intellect Environments Hub, application-only bounty)
Price
Figma: $2,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Intercom (Prime Intellect Environments Hub, application-only bounty)
Price
Intercom: $1,200
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Mind2Web web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
Mind2Web: $2,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MLE-Bench (OpenAI) reimplemented as an environment: ML engineering competitions for an agent (Prime Intellect Environments Hub, application-only bounty)
Price
$1,200
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Monday.com (Prime Intellect Environments Hub, application-only bounty)
Price
Monday: $1,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
nanoGPT speedrunning environment: an agent trains a small language model as fast as it can (Prime Intellect Environments Hub, application-only bounty)
Price
$2,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Vercel's Next.js eval suite implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
Next.js Bench: $3,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Notion (Prime Intellect Environments Hub, application-only bounty)
Price
Notion: $1,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
OpenHands agent framework integration with the Environments Hub (Prime Intellect Environments Hub, application-only bounty)
Price
$2,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
OSWorld computer-use benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
OSWorld: $1,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
PaperBench (OpenAI) reimplemented as an environment: an agent replicates research papers (Prime Intellect Environments Hub, application-only bounty)
Price
$1,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Software-library environment for agent tasks in PyTorch (Prime Intellect Environments Hub, application-only bounty)
Price
PyTorch: $5,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Shop-R1 reimplemented as an environment: an agent simulating online shopping behaviour (Prime Intellect Environments Hub, application-only bounty)
Price
$2,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Stripe (Prime Intellect Environments Hub, application-only bounty)
Price
Stripe: $1,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
SWE-bench Verified coding benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
SWE-bench Verified: $3,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
TAU2 (Sierra) extended training set generation pipeline: customer-service tool-use tasks generated for training (Prime Intellect Environments Hub, application-only bounty)
Price
$1,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment for Prime Intellect's Environments Hub (programme announced 2025-10-27, now closed). One programme: 18 of its rows were filed on 2026-10-14; the Application-Only tab holds about 85 priced rows from $500 to $5,000, and these rows add categories not yet on the board. Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
WebArena web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
WebArena: $2,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Webflow (Prime Intellect Environments Hub, application-only bounty)
Price
WebFlow: $2,500
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
WebVoyager web-agent benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
Price
WebVoyager: $1,200
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
MCP backend simulator and agent environment for Zapier (Prime Intellect Environments Hub, application-only bounty)
Price
Zapier: $2,000
per environment (bounty)
When
Oct 2025
What to know
Bounty paid on completion to the builder of an open-source environment published on Prime Intellect's hub (program announced 2025-10-27, now closed). Open and non-exclusive, so a floor under private vendor prices.
Source
Prime Intellect Environments Program bounty sheet, Application-Only tab
The payer's own page
Anthropic's discussed spend on RL environments bought from vendors over the following year
Price
more than $1 billion
total over the next year
When
Sep 2025
What to know
TechCrunch citing The Information; a discussed budget, not a signed contract. Read with the $200 to $2,000 per task band: at those prices it is roughly 0.5 to 5 million tasks.
Source
TechCrunch, Silicon Valley bets big on environments to train AI agents
Press report
Software engineer salary at Mechanize for building RL environments (2025 report)
Price
$500,000 salaries
per year (salary)
When
Sep 2025
What to know
Employee salary, not a contractor rate; TechCrunch contrasts it with hourly contractor pay at Scale AI or Surge.
Source
TechCrunch, Silicon Valley bets big on environments to train AI agents
Press report
Recommended spend per RL task for frontier post-training (Mechanize analysis)
Price
~$500 per RL task now, rising roughly fivefold to a few thousand dollars within a year
per RL task
When
Aug 2025
What to know
An argument from compute cost by a company that sells RL environments, not an observed price: about $480 of compute per task per training run, and $2,400 over a task's life at five reuses.
Source
Mechanize
A vendor's published price guide
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)
Price
$2,400 for each RL task
per RL task, compute over its life
When
Aug 2025
What to know
A modelled compute cost a lab bears when it trains on one task, not a price paid for the task. Written by a company that sells RL environments, to argue labs should spend at least about $500 per task.
Source
Mechanize, Cheap RL tasks will waste compute
A vendor's published price guide
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)
Price
£999,600
per contract (Jan to Apr 2025)
When
Jun 2025
What to know
A government buyer's published award for an environment made to its order: the requirement names one complex cyber range environment plus its tasks, solvers, graders and human baselines, so this is the nearest public figure to a per-environment price for a high-fidelity private environment. Awarded 2025-01-10 by call-off from a dynamic purchasing system, contract 2025-03-13 to 2025-04-30.
Source
UK Contracts Finder (official API record), DSIT award notice: WP 1: Cyber Range Scenarios, Enterprise Networks
The payer's own page
Novel research-level maths problem with solution written for the FrontierMath benchmark
Price
up to $1,000 apiece
per problem
When
Jun 2025
What to know
Paid by Epoch AI to professional mathematicians; problems were graded into three difficulty tiers.
Source
Financial Times (Anjana Ahuja, republished by LIMS)
Press report
Synthetic multimodal web-agent trajectory from exploration (Explorer, 94K successful trajectories)
Price
28 cents per successful trajectory
per successful trajectory
When
May 2025
What to know
Producer's own compute and API cost, no humans; over 94K successful trajectories were produced.
Source
arXiv, Explorer (v4)
The payer's own page
Human expert baseline attempt on an 8-hour AI R&D task (METR RE-Bench)
Price
approximately $1,855
per expert attempt (8-hour task)
When
May 2025
What to know
Average paid per expert run; over the 8-hour budget this is about $232 an hour. An agent run of the same length cost about $123 in tokens.
Source
arXiv, RE-Bench (v2)
The payer's own page
150,000 synthetic web-agent trajectories with judge annotations, generated on 1,200 V100 GPU hours (InSTA)
Price
costs $521.55
per dataset of 150k trajectories
When
May 2025
What to know
Open models (Qwen3) on spot GPUs, no human annotation; the authors call it a fraction of what industry labs spend on agents.
Source
arXiv, InSTA (v2)
The payer's own page
One 12-hour agent rollout replicating a research paper (o1 IterativeAgent on PaperBench)
Price
$400 in API credits
per rollout on one paper
When
Apr 2025
What to know
Inference cost of one long-horizon trajectory, plus about $66 to grade it; it shows why labs pay for tasks where a long rollout teaches something.
Source
arXiv, PaperBench (v3)
The payer's own page
Human baseliner timing software and research tasks for METR's HCAST benchmark
Price
between $50 - $100 per hour, plus $25 - $150 per hour in performance bonuses
per hour (base, before bonuses)
When
Mar 2025
What to know
With bonuses for fast, successful completion the total runs $75 to $250 an hour.
Source
arXiv, HCAST
The payer's own page
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)
Price
£590,500
per contract (Jan to Jun 2025)
When
Mar 2025
What to know
Sister work package to the SpecOps enterprise-network range: the same buyer paid a second vendor for critical-infrastructure range scenarios. The notice does not state how many scenarios, so no per-environment figure. Awarded 2024-12-12, contract 2025-01-02 to 2025-06-30.
Source
UK Contracts Finder (official API record), DSIT award notice: Cyber Evaluations, WP1, Cyber Range Scenarios
The payer's own page
Synthetic GUI/web agent trajectory generated by replaying web tutorials (AgentTrek)
Price
$0.551 per trajectory
per trajectory
When
Mar 2025
What to know
Model API cost with no human annotators; the abstract rounds it to $0.55. Executing 1,000 tasks with GPT-4o cost about $215 in the same paper.
Source
arXiv, AgentTrek (v2)
The payer's own page
Real freelance software tasks from Upwork used as the SWE-Lancer eval set
Price
$50 bug fixes to $32,000 feature implementations; $1 million total across 1,400+ tasks
per task (original freelance payout)
When
Feb 2025
What to know
These are what clients originally paid freelancers for each task, used by OpenAI to value the eval; roughly $700 per task on average (derived).
Source
OpenAI (SWE-Lancer paper, arXiv)
The payer's own page
Human-annotated web-agent trajectory in the Mind2Web training set (as costed by the Synatra authors)
Price
each trajectory costing $0.8
per trajectory
When
Nov 2024
What to know
A third party's costing of an academic crowd-annotated dataset; the per-trajectory figure is the academic floor for human computer-use demonstrations.
Source
arXiv, Synatra (v2), comparison with Mind2Web
The payer's own page
Synthetic web-agent demonstration converted from tutorials and indirect knowledge (Synatra)
Price
$0.031 each
per demonstration
When
Nov 2024
What to know
The authors put it at 3% of the cost of a human demonstration and found it can beat an equal number of human demonstrations from limited domains.
Source
arXiv, Synatra (v2)
The payer's own page
Development time on an evaluation task for the UK AI Safety Institute (evals bounty indication)
Price
£100-200 per development hour
per development hour
When
Nov 2024
What to know
The rate the buyer uses to size a task bounty, paid only for tasks that succeed; stage-two applicants also get a GBP 2,000 award toward compute.
Source
UK AI Security Institute, evals bounty (blog)
The payer's own page
One successful dangerous-capability or autonomous-agent evaluation task built for the UK AI Safety Institute (evals bounty)
Price
£3000-£15,000 for a successful task
per successful task
When
Nov 2024
What to know
A government evaluator buying tasks directly from builders (individuals or organisations), paid in full only once the evaluation meets its criteria; exceptional work may earn more.
Source
UK AI Security Institute, evals bounty (blog)
The payer's own page
Private agent-evaluation task for METR, priced by the hours a professional would need to complete it
Price
$300 per hour it would take a professional to complete the task
per hour of professional time the task takes to complete
When
Nov 2024
What to know
The later rule of METR's task bounty (paused to new submissions on 2024-11-18, over $100,000 paid out): a 10-hour task would earn $3,000. Per-task value depends on task length, which varies, so no per-task figure is derived.
Source
METR Task Development Guide, Introduction
The payer's own page
Expert-written benchmark question for Humanity's Last Exam (eval set)
Price
$5,000 each for the top 50 questions, $500 each for the next 500
per accepted question
When
Sep 2024
What to know
Prize money from a $500,000 pool paid by the eval's builders (CAIS and Scale AI) for questions frontier models could not answer.
Source
Center for AI Safety
The payer's own page
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)
Price
£440,000
per contract (Jun to Oct 2024)
When
Aug 2024
What to know
Four-month build of private agent scenarios; scenario count not stated. The supplier is named 'Grey Swan Security Inc'; the agent-safeguard benchmark work matches Gray Swan AI's line of business, but the notice alone does not confirm it is the same company.
Source
UK Contracts Finder (official API record), DSIT award notice: offensive cyber capability and agent safeguards measurement
The payer's own page
Suites of tasks ('evals') for evaluating AI systems, created to order (Rethink Priorities for the UK AI Safety Institute)
Price
£232,517.64 (first awarded at £155,011.76)
per contract (Jan to Mar 2024)
When
Feb 2024
What to know
Twelve-week task-suite build, extended in time and value by half again (a sign that a first task order grows once the buyer sees the tasks). Task count not stated.
Source
UK Contracts Finder (official API record), DSIT award notice: Creating suites of tasks ('evals') to use in the evaluation of AI systems
The payer's own page
One private agent-evaluation task implementation for METR that takes an experienced engineer about 10 hours (worked example)
Price
the base price would be $3000 ... pay $4500 overall
per task implementation
When
Dec 2023
What to know
METR, an evaluator that uses the tasks itself, pays the builder directly with no vendor in between: a base of 3x the human cost to solve the task, plus up to a 50% bonus. Tasks must stay off the public web and out of training data.
Source
METR, Bounty: Diverse hard tasks for LLM agents
The payer's own page
Crowd annotator demonstrating web tasks for the Mind2Web agent dataset
Price
estimated hourly rate of $10.10
per hour
When
Dec 2023
What to know
Academic rate set to Ohio minimum wage; the human end of the trajectory market before labs paid experts.
Source
arXiv, Mind2Web (v3)
The payer's own page
PhD-level expert writing and validating GPQA benchmark questions
Price
approximately $95 per hour
per hour (average; maximum possible $150)
When
Nov 2023
What to know
Paid as a $10 base per question plus bonuses when experts agree and non-experts fail, so pay tracks question difficulty.
Source
arXiv, GPQA
The payer's own page

What yours might fetch

A first answer or two gives a range. The full valuation asks the rest and shows where every figure comes from.

What is your data worth?

Kind of data

How many

$17,000to$1,100,000

On one answer, against 3 public prices. An indication, not an offer.

No data upload
No account
Free

Why labs pay for it

Labs train agents by having them attempt tasks thousands of times with a grader scoring each try. Good tasks are hard to write and expire once models pass them, so labs buy them continuously.

RisingFrontier labs sign six- and seven-figure contracts a quarter for environments, and some buy tens of thousands of tasks a month (2026). Generic web-app replicas are commoditising.

Who holds it

Independent builders and small studios
Research groups
Companies with internal evaluation sets
Domain experts

What it looks like

Tasks with an automatic grader for each
A working replica of an app for an agent to practise in
A private evaluation set no model has seen
A curated dataset in a narrow field

Raises the price

Graders that cannot be fooled
A professional domain where the builder has real expertise
Difficulty calibrated so models pass sometimes, not always
Exclusivity, which buyers pay several times more for

Lowers it

Tasks a model already passes
Graders that reward shortcuts
Replicas of common apps, which many vendors now build

The rights question

You must own the environment and everything inside it. Tasks derived from someone else's proprietary material carry that material's restrictions.

How we take it to buyers

We take finished, tested work to labs and environment buyers, and broker buyers' requests to builders who can meet them.

Who buys it

Prime Intellect

$100 to $500 for open bounties; $1,000 to $5,000+ for application-only tasks.

Prime Intellect, environments programme

Sieve

Pays after delivery is checked ("Sieve Checks them, then pays"); recording is paid per hour and no rate is published.

Sieve, partners page

Datarade

The seller sets the price. Datarade takes 30% commission, or 20% to 15% on a paid plan of $6,000 to $12,000 a year.

Datarade, provider page

DataVendor (HUD)

Publishes nothing about price.

DataVendor, site

Grepped

Its own pages cite $20K to $5M per company; its estimator shows a sample of $29K to $41.5K.

Grepped, business page

FileYield

No figures. It charges sellers 8% and buyers 6% commission.

FileYield, sell data page

Opendatabay

The seller keeps 70% of a sale up to $100, rising to 95% above $250,000 (the commission falls from 30% to 5% as the sale grows). In a brokered sale the seller gets the agreed provider price and Opendatabay sets its own price to the buyer.

Opendatabay, fees and commission schedule

Every buyer and what it publishes

Where datasets and environments comes from

The whole supply map

more than 3,500 products from over 300 providers
AWS Data Exchange lists more than 3,500 data products from over 300 providers, and over 1,000 of them are free.
Reaches buyers: Cloud marketplace
AWS, Why AWS Data Exchange, 11 Oct 2026
1 751 247 datasets
The European data portal lists 1,751,247 datasets from 212 catalogues in 36 countries.
Reaches buyers: Open download
data.europa.eu, The portal in numbers, 16 Oct 2026
586,902
Data.gov lists 586,902 datasets.
Reaches buyers: Open download
Data.gov, Datasets catalog, 16 Oct 2026
2700+ trusted data providers
Datarade lists data products from more than 2,700 data providers across more than 560 data categories.
Reaches buyers: Broker marketplace
Datarade, The #1 Global Data Marketplace, 11 Oct 2026
819 datasets
Defined.ai's dataset marketplace lists 819 off-the-shelf datasets. Its homepage cites a contributor crowd of 1.6M+ across 150+ countries.
Reaches buyers: Marketplace plus contributor crowd
Defined.ai, Datasets, 11 Oct 2026
1,092,755
The Hugging Face Hub lists 1,092,755 datasets, almost all free to download.
Reaches buyers: Open download (hub)
Hugging Face, Datasets, 16 Oct 2026

What a buyer's counsel will ask

19 questions apply to datasets and environments, and 12 of them block a sale until settled. Most data that fails to sell fails here, not on price.

Run the rights check Read every question

Other kinds of data

Company records Source code Video and footage Audio and speech Images Text and archives Physical-world data Expert and human work

Hold datasets and environments? Describe it and we take it to the buyers.

Offer datasets and environments I need this data

In this topic

What sells

Sell datasets and environments to AI labs: prices, buyers and rights | AIData.Broker