Datasets, tasks and environments · Graded coding tasks · Updated 30 Oct 2026

Who buys graded coding tasks for AI training, and what does one task pay?

AIData.Broker represents the companies and people that hold graded coding tasks when they sell to AI labs: we value what they hold against public prices, describe it without their name to every buyer that takes it at the same time and help negotiate the terms. Our fee is 12% of the first $250,000 a sale pays, and nothing if no sale closes.

Software tasks with a grader that runs: a repository, an issue or a goal, and tests that fail before the work and pass after it. Labs train and test coding agents on them, and the public benchmarks have leaked into training.

Value your graded coding tasksHow to Sell a Dataset: Where to Sell It and What It's Worth

No data upload
About two minutes
No email needed

Prices on record: 25
asking prices off the shelf, made to order, cost to produce, paid to the people who make it
Who holds it: Teams with their own evaluation suites, Coding-assessment and interview platforms, Coding courses with graded exercises
Who has paid, offered or listed: Getly, SemiAnalysis, Epoch AI, Snorkel AI, Mechanize

The short answer · figures from the market's records, last reviewed 30 Oct 2026

Who buys graded coding tasks for AI training, and what does one task pay?

Frontier labs, mostly through the data and environment companies that sell to them. Analysts report labs paying $5,000 or more for one decent RL coding task, and Epoch AI's interviewees put a rare, very complex software-engineering task with its grader at $20k. What vendors pay the engineers who write them sets a floor: Scale AI's Outlier pays up to $1,800 for a terminal task that stumps frontier models. A task a model generated costs cents to make ($0.00613 a task in one paper), so buyers pay for tasks people wrote and checked, with a grader that runs.

What buyers have paid, offered or asked for graded coding tasks

The whole price board

Every price in the market's records about graded coding tasks, each with its source and the day it was read. Open a row for the source's own words.

25 public prices

551 code-agent trajectories marked PASS by an LLM judge (JSONL with reasoning chains, tool calls and final answers), commercial use allowed, no resale of the file, produced by an autonomous agent in a sandboxed Python environment
A retail asking price from a seller with five listings on Getly since September 2026. The check is a language-model judge, not an environment's own verifier, and the data is machine-generated, so it prices cheap synthetic trajectories, not the human-verified trajectories labs buy.
Datasets, tasks and environmentsAsking prices off the shelfSep 2026A catalogue or marketplace listing
Getly marketplace, Code Agent Trajectories: 551 Verified PASS (Jurgen's Data Lab)
$100
per file of 551 trajectories
$0.18 per demonstration
One decent RL coding task bought by a frontier lab from a data or environment company
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.
Datasets, tasks and environmentsMade to orderJul 2026Press report
SemiAnalysis, The Future of Meta Superintelligence: A 1 Year Progress Update (Kan, Martin-Prin, Eliahou Ontiveros, Patel)
$5,000
per coding task (a floor: 'plus')
One especially complex software-engineering RL task with its grader (upper end reported to Epoch AI)
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.
Datasets, tasks and environmentsMade to orderJan 2026Press report
Epoch AI, An FAQ on reinforcement learning environments (Denain and Barber)
$20K
per task
Snorkel AI's Open Benchmarks Grants programme, funding independent builders of agent benchmarks and their task pipelines (Terminal-Bench, TB-Science, ARC-AGI-3, OSWorld 2.0 and others), expanded from $3M to $30M
A data vendor's own budget for paying outside teams to build agent benchmarks and environments, so a ceiling on one funder's demand rather than a price. Grants include a red-team programme for reward-hacking and broken verifiers and a research fellowship. Snorkel also sells expert data to labs, so the programme is partly a market-making cost.
Datasets, tasks and environmentsCost to produceOct 2026The payer's own page
Snorkel AI, Snorkel's $30M commitment to Open Benchmarks (Vincent Sunn Chen, 7 Oct 2026)
$30M
per programme, total across grants
Machine-synthesised long-horizon terminal-agent task with environment, reference solution and verifier, accepted after validation (Recursive Synthesis for Long-Horizon Terminal Tasks)
Cost of generation by model with no human authoring, 37,484 tasks after fifteen rounds from 639 seed tasks. The same paper says human-authored long-horizon tasks are "often costing hundreds to thousands of dollars per task" (citing earlier work), the range the paid task rows in this store measure.
Datasets, tasks and environmentsCost to produceAug 2026The payer's own page
arXiv 2608.05466, Recursive Synthesis for Long-Horizon Terminal Tasks (v3, 12 Aug 2026)
About $0.05
per verified synthesised terminal task
Labour cost of one good RL coding task at Mechanize: an engineer paid $400k+ a year expected to make one good task a week
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.
Datasets, tasks and environmentsCost to produceJul 2026Press report
SemiAnalysis, The Future of Meta Superintelligence: A 1 Year Progress Update
$400K
per engineer-year, at one good task per week
$7,692 per task
Machine-built executable terminal-agent training environment (Terminal-World, environment building by Gemini-3-Flash)
Model API cost of the environment-building stage only, with an 83.1% construction success rate; task generation is a separate $0.015 per task. A floor under what a buyer pays a vendor for an environment, not a price.
Datasets, tasks and environmentsCost to produceMay 2026The payer's own page
arXiv 2605.20876, Terminal-World: Scaling Terminal-Agent Environments via Agent Skills, Table 5
$0.083
per generated terminal environment (API cost)
Machine-made terminal-agent trajectory including task, environment and teacher rollout (Terminal-World, DeepSeek-V3.2 and Gemini-3-Flash)
API cost of task generation, environment building and trajectory collection for 5,723 trajectories; no human authoring. Compare Getly's asking $0.18 to $0.85 per verified trajectory (held).
Datasets, tasks and environmentsCost to produceMay 2026The payer's own page
arXiv 2605.20876, Terminal-World: Scaling Terminal-Agent Environments via Agent Skills
$0.17
per trajectory (full pipeline API cost)
Environment generated automatically by coding agents from a theme (rules, observations, rewards and levels), 36 heterogeneous environments built
Producer's own model cost, no humans, and the overall success rate of the generation pipeline was 65.0 percent (80.0 percent with human review of themes). These are rule-based game-like worlds, not replicas of real software, so this is the floor of the cost range that vendors' $20k to $300k replicas sit above.
Datasets, tasks and environmentsCost to produceNov 2025The payer's own page
arXiv 2511.19304, AutoEnv: Automated Environments for Measuring Cross-Environment Agent Learning
$4.12
per environment, API cost of generation
Software-engineering task instance created by having a model rewrite a function to introduce a bug in a repository, then validated by the repository's tests (SWE-smith, pallets/flask: 402 candidates, 267 qualified)
The paper sets this against collecting 11 SWE-bench-style instances for the same repository, an estimated 38 hours of human labor (207.27 minutes each). The figure excludes the cost of building the repository's execution environment (10 minutes to define installation and tests), so it prices the task, not the environment, and it is the producer's own cost, not a price paid.
Datasets, tasks and environmentsCost to produceApr 2025The payer's own page
arXiv 2504.21798, SWE-smith: Scaling Data for Software Engineering Agents (appendix, cost comparison)
$0.0061
per task instance, API cost
Machine learning practitioners in Mercor's talent network, standing listing titled 'Technical Reviewer - RL Environment Terminal Benchmarking (Agentic AI)': writing modeling problems with reference answers and grading model-written training and evaluation code (remote independent contractor, part-time, paid weekly)
A standing listing for a pool, not one project, and the range is for 'comparable machine learning projects', set per project by scope and depth; the body describes problem writing and grading rather than building an environment, so treat it as the pay band of the vendor's reviewers. A second listing with the identical text sits at list_AAABmWu3eSKR9nSVhWNLZajY. The listing no longer accepts applications.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Mercor Jobs, Technical Reviewer - RL Environment Terminal Benchmarking (Agentic AI)
$70 to $120
per hour
Terac: software engineers designing realistic full-stack coding tasks and tests used to train and evaluate frontier AI coding agents (SWE coding benchmarks and RL environments)
Independent contractor, 10 to 20+ hours a week, paid weekly after a screening interview and a paid work trial. The page header shows $75 per hour and the body gives the $45-75 range.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Terac opportunity, SWE Coding Benchmarks & RL Environments
$45 to $75
per hour
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
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.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Outlier AI (with Scale AI), Terminal-Bench expert page
Up to $1,800
per task, typically 6 to 7 hours
Recorded AI-coding-assistant session (an agent trajectory of about 50 turns) contributed to a lab via Mercor
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 to $200; the page header now reads "$200 / task". Closed: Mercor shows the posting no longer accepting applications.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Mercor, Builders & Developers: Share Your AI Coding Sessions
$200
per accepted session (about 50 turns, 2 to 3 hours)
Software engineer at Mechanize, hired to build environments and evals for frontier coding agents
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.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Mechanize, Apply page
$300K to $400K
per year (base salary, plus equity and bonus)
Terminal-style coding task that stumps frontier models (Outlier, run with Scale AI)
Contributor payout ceiling, not the lab's price; the page says rates vary by expertise and lower rates apply to non-core work.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Outlier (Scale AI)
Up to $1,350
per task (paid to the contributor)
Reproducible physical-science task an AI agent must complete in a Linux terminal, Turing Physical Sciences Expert
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.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
AITraining.jobs, Turing Physical Sciences Expert, AI Training (Turing's posting, mirrored)
About $300
per approved task of six to eight hours
$43 per hour of work
Software engineers reviewing and verifying coding environments, tasks and evaluation harnesses used to benchmark AI agents (Terac, remote independent contractor, screening interview first)
What an expert network pays an engineer to review realistic programming tasks and their evaluation harnesses for an AI developer's agent benchmarks (verification, not authoring from scratch; the client is not named). Listing created 2 Oct 2026; sister Terac programmes pay $40 to $90 an hour for the same kind of work. Hourly pay only, no per-environment price.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Terac opportunity page, Software Engineers: Paid Code Review for AI Agent Evaluation
$65
per hour
Software engineers in South Asia, and in Brazil and Argentina, verifying coding tasks and programming environments built to evaluate AI agents (Terac, remote independent contractor)
The same programme at the same $40 an hour is posted for Brazil and Argentina (https://terac.com/opportunities/6iC1ggB7S3drHEp8), so it is one observation. Terac's review roles without a regional restriction pay $65 and $75 an hour. The client building the suite of programming environments is not named.
Datasets, tasks and environmentsPaid to the people who make itOct 2026The payer's own page
Terac opportunity page, South Asian Software Engineers: Coding Tasks for AI Evaluation
$40
per hour
Terminal-based scientific task for Terminal Bench Science (environment, expert solution, automated tests), Turing Scientific Computing Expert
Five-week contract at 40 hours a week; Outlier pays up to $1,350 for a comparable Terminal-Bench task (already on the board).
Datasets, tasks and environmentsPaid to the people who make itSep 2026The payer's own page
AITraining.jobs, Turing Scientific Computing Expert, Physical Sciences (Turing's posting, mirrored)
$300
per approved task
Auditor of repository-level software-engineering benchmark tasks (reference patches, test harnesses) for Mercor
QA labour on verifiable coding tasks, US only; page said posted a month before reading.
Datasets, tasks and environmentsPaid to the people who make itSep 2026The payer's own page
Mercor, SWE-Bench Task Auditor
$70 to $90
per hour
Engineer building the pipeline that captures and scrubs coding-agent (Copilot CLI) traces for a lab dataset (Turing contract)
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.
Datasets, tasks and environmentsPaid to the people who make itSep 2026A catalogue or marketplace listing
aitrainer.work, Turing: Software / Platform / Data Engineer, Agent Trace Collection
$100
per hour (3-month contract, 20+ hours a week)
Software engineering intern at Mechanize designing RL tasks (prompt, environment and automated grader)
Intern rate, already above most expert contractor rates for environment work; each task is a self-contained software challenge with its own grader.
Datasets, tasks and environmentsPaid to the people who make itNov 2025The payer's own page
Mechanize on Ashby, Software Engineering Intern
$150
per hour
SWE-bench Verified coding benchmark implemented as an environment (Prime Intellect Environments Hub, application-only bounty)
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.
Datasets, tasks and environmentsPaid to the people who make itOct 2025The payer's own page
Prime Intellect Environments Program bounty sheet, Application-Only tab
$3,500
per environment (bounty)
Real freelance software tasks from Upwork used as the SWE-Lancer eval set
These are what clients originally paid freelancers for each task, used by OpenAI to value the eval; roughly $700 per task on average (derived).
Datasets, tasks and environmentsPaid to the people who make itFeb 2025The payer's own page
OpenAI (SWE-Lancer paper, arXiv)
$50 to $32K
per task (original freelance payout)

Where this data comes from, and who lists it

The whole supply map

$1,000 to $2,000 per environment bounty; trajectories by request
Chakra Labs paid outside developers fixed bounties of $1,000 to $2,000 per app clone for its Dojo environments (SAP, QuickBooks, ZenDesk, HubSpot look-alikes), then stopped maintaining the public repository because Dojo moved to an enterprise product. Its Pango computer-use trajectory dataset on Hugging Face (16.3 GB) is available only by contacting the company, with no price shown.
Reaches buyers: Environment vendor: bounties to builders, enterprise sales to labs and companies
Chakra-Network/dojo-spas README, 21 Oct 2026
88 tasks
Halluminate's August 2026 benchmark has 88 tasks based on anonymized private-equity transactions, written and reviewed by practicing deal professionals; seven frontier models averaged at most 51 percent. The company says it turns such failure modes into reinforcement-learning environments for frontier labs.
Reaches buyers: Expert-written tasks that become RL environments for labs
Fortune, Exclusive: Nine-person Halluminate raises $30 million, counts four top U.S. AI labs as customers (Wen Shao), 1 Oct 2026
the majority of world-model training material (its chief executive's belief)
Worldmodeldata's CEO believes video game data will make up the majority of world-model training material, later tuned with task-specific data.
Reaches buyers: CEO forecast
WIRED, The Next Evolution of AI Is Learning From Your Dodgy Gaming Skills (2026-09-28), 28 Sep 2026
one program, value not stated
A listed data vendor reports a second program with its largest customer that builds reinforcement-learning environments for computer-use agentic tasks, with no contract value, number of environments or price stated; the company reports quarterly revenue of $92.1 million, up 58 percent.
Reaches buyers: Environment vendors selling to frontier labs
Innodata Inc., Form 8-K Exhibit 99.1, second quarter 2026 results (6 Aug 2026), 6 Aug 2026
~365,000 tasks in 23 tasksets
Prime Intellect ships first-party integrations for 23 published agentic tasksets, about 365,000 tasks in all (about 198,000 software engineering, 28,600 terminal and 137,600 search), re-uploaded and validated where the originals needed fixing. These are public research datasets, so the open supply of agent tasks is far larger than any private set.
Reaches buyers: Open tasksets run on Prime Intellect infrastructure
Prime Intellect, Scaling Agentic RL: 365,000+ Environments for SWE, Terminal, and Search, 22 Jul 2026
400 tasks, over 2,000 expert hours
$OneMillion-Bench, a benchmark of 400 open-ended expert tasks in finance, law, healthcare, natural science and industry, took over 2,000 expert hours to curate (more than five hours a task); each task is valued at a senior professional's estimated time at the market hourly wage, and all together they are valued above $1 million.
Reaches buyers: Expert-curated agent benchmark tasks
$OneMillion-Bench: How Far are Language Agents from Human Experts? (arXiv 2603.07980), Mar 2026

Raises the price

A grader that runs: tests that fail before and pass after
Tasks today's frontier models fail
Written and checked by people, with the time each took
Containerised: the environment builds and runs anywhere

The rights question

Tasks written from a company's own code are the company's to license; tasks written for a client or a vendor under contract are theirs. A task a model drafted has to say so, because code buyers now set rules on AI-written work, and tasks already public are worth little once they have leaked into training.

The guide on what else to settle