How to Sell an RL Environment or Task Set to AI Labs

An RL environment or task set reaches an AI lab through one of five routes: a build contract with the lab, a marketplace listing, a sale to a human-data company or larger builder, a paid task bounty, or unpaid hosting on an open hub. The route decides who names the price, who keeps the environment and when money moves.

Buyer evidence is thin and mostly stated by sellers, prices are claims that disagree, and not selling is a valid outcome. Selling an environment is one corner of AI training data licensing, and DataPacts takes no cut of any sale. Every figure here is a claim by the party named, with its date; the method is explained under how we source, date and check every figure.

What a Lab Receives When It Buys an Environment or Task Set

A lab buys one of five units: a task set with its graders, an environment sold together with the task set that runs on it, a custom build delivered under contract, the trajectories agents leave in an environment, or a codebase that seeds new tasks. The table compares the five units by what the buyer receives, with one example on record each, as of 2026-11-10.

The components of an environment are defined on the hub, where what RL environments are and who buys them is covered in full. The unit decides the route, because DataVendor, a marketplace, lists an environment only together with the taskset that runs on it (checked 2026-11-10).

Five Routes to Sell an Environment or Task Set

The route decides who sets the price, who keeps the environment and when money moves, more than the environment itself does. The table orders the five routes from most to least seller control over price, as of 2026-11-10, and every figure is a claim of the named party.

A Direct Build Contract With a Lab

A direct contract is a commissioned build: the lab orders tasks or an environment, and contract size, not unit price, sets the deal. Epoch AI's interviewees (published 2026-01-12) describe contracts of often six to seven figures per quarter, and one neolab researcher put a contract at USD 300,000 to 500,000 with no period stated. Anthropic's S-1 (2026-10-26) says its training data includes datasets licensed or commissioned from third parties, but no lab's own buying terms for environments are on record.

Proximal's master services agreement (last updated 2026-07-06) shows a builder-side template: invoices run monthly and undisputed amounts are due within 30 days (checked 2026-11-09). Under the same agreement Proximal owns the deliverables, so the customer receives an internal-use license for the term.

Listing an Environment for Sale on a Marketplace

DataVendor, HUD's marketplace, is a public venue on record where an organization can list a taskset with its environment, set its own price and sell to buyers invited from a waitlist. DataVendor states that an environment "is sold with the HUD taskset that runs on it; it cannot be listed on its own" (checked 2026-11-10). Onboarding is self-serve for organizations only (checked 2026-11-09).

DataVendor states that vendors set the price, that it takes 20% of earnings at withdrawal, and that the minimum withdrawal is USD 1,000 before fees (checked 2026-11-09). It lists and hosts the assets and does not itself acquire them. See also DataVendor's seller terms in full.

Selling to Human-Data Companies and Larger Builders

Human-data companies buy RL-environment work in three ways: they license trajectories, build environments on a holder's records, or acquire the builder outright.

• Mercor: Mercor.io Corporation acquired Sepal AI, Inc. (Orrick, 2026-02-06; price undisclosed), and Mercor announced on 2026-07-09 that it will acquire Deeptune, with value undisclosed and closing not confirmed as of 2026-11-09.
• micro1: it states "We license qualifying trajectories for $1-2M annually, with your approval" (checked 2026-11-09).
• Turing: Project Lazarus pays a revenue share with nothing upfront and the split is not published (checked 2026-11-09).

These counterparties may resell what they acquire, so the builder's buyer is not always the end lab. The directory of RL environment companies lists builders that also buy.

Task Bounties and Open Hubs That Pay Little or Nothing

Bounties pay per accepted task in the tens to hundreds of US dollars, and open hubs host environments without paying their creators. Prime Intellect's program (blog of 2025-10-27) lists Open Access bounties "in the $100-500 range", awarded on completion. Refresh states USD 30 to 75 per task, with the price agreed upfront and paid on approval, and Shipd shows bounties of USD 80 to 220 per item (checked 2026-11-09).

General Reasoning's OpenReward deploys from a Dockerfile and a GitHub repository, states no payment to creators and says "Your environment code stays on GitHub" (checked 2026-11-09). BenchFlow's PostTrain Arena states no compensation and its repository is licensed AGPL-3.0, while Prime Intellect's verifiers library is MIT.

What Labs Test Before Buying: Difficulty, Grading and Build Checks

Labs and marketplaces check three things before they pay: that tasks are hard enough to teach, that the grader cannot be gamed, and that the environment builds and runs cleanly. The table lists each check with its stated threshold, as of 2026-11-10, and the party that states it.

Difficulty thresholds come from interviewees and builders, not from a lab, and Mechanize states (2025-08-18) that a minimum efficient spend is about USD 500 per task, with compute near USD 480 per run and USD 2,400 per task over five reuses, so on its reasoning a cheap task costs the buyer more in compute than it is worth.

Grader design and how reward hacking is caught are covered under verifiers, graders and rubrics.

Preparing the Sale Step by Step

Preparation runs in five steps, and the first three decide whether any route accepts the environment at all.

Package the Tasks, Graders and Build

A sellable package holds the tasks, a grader for each task, a reproducible build and baseline results from more than one agent. DataVendor requires "A successful build" for environments, and Abundant's stress test spans at least five runs across at least three agents (2026-11-03).

Clear Every Input Before Listing

Every input inside a task set must be the seller's to license: contributor work, copied software and data, and outputs of commercial models. Shipd's terms (s. 3.1, checked 2026-11-09) have the contributor assign all work product to Shipd, which shows how platforms secure contributor work. Anthropic's Commercial Terms (s. D.4, effective 2025-06-17) bar using the services to train competing AI models, and the Gemini API terms (effective 2026-03-23) bar using them to develop competing models, a restriction to check where tasks or reference solutions were generated with them. The rules on rights in an RL environment follow input by input.

Strip Credentials and Personal Data From Tasks

Removing credentials and personal data is the seller's job, because DataVendor, the one marketplace on record, cleans purchased repositories and raw data but not tasksets and environments.

• Before listing: DataVendor's mandatory free checks scan the listing text for credentials and personal data such as emails or phone numbers (checked 2026-11-10).
• After purchase: DataVendor cleans purchased files (repositories, raw data, Harbor tasks) of personal information within 2 business days, and tasksets and environments are not cleaned (checked 2026-11-09).

Replicas of real software can carry customer records, so the seller screens task seed data too. The method is set out under PII removal and de-identification before licensing.

Set the License Scope: Who Keeps the Environment

The license decides whether the seller can sell the same environment again, and builders' published documents point in three directions.

• Seller keeps ownership: DataVendor states that a sale licenses the assets for a commercial use rather than transferring them, that sales are final with no refunds, and that the buyer's entitlement survives edits or unpublishing (checked 2026-11-09).
• Builder keeps deliverables: Proximal's agreement (s. 8.1 to 8.3, last updated 2026-07-06) says Proximal owns the deliverables, may reuse them for other customers absent customer confidential information, and retains learnings including model training.
• Contributor assigns everything: Shipd's terms (s. 3.1) assign all work product to Shipd (checked 2026-11-09).

Two founders told Epoch AI (2026-01-12) that exclusivity runs roughly 4 to 5 times the non-exclusive price, a claim and not a market figure. The terms of exclusive vs non-exclusive data licenses and their lock-in are covered separately.

Quote, Invoice and Get Paid

No price list exists for RL tasks, so the quote rests on claims that disagree: USD 200 to 2,000 per task from Epoch AI's interviewees (2026-01-12) against Mechanize's floor of about USD 500 per task (2025-08-18). DataVendor states that opportunities are priced around USD 500 to 1,500 a task, with hundreds to thousands of tasks per request (checked 2026-11-10). SemiAnalysis via Epoch AI puts a UI gym at about USD 20,000, and one interviewee puts a Slack-like replica at around USD 300,000 (2026-01-12).

Payment triggers differ by route. Refresh pays on approval, Prime Intellect awards bounties on completion, DataVendor earnings are available as soon as the buyer pays with a withdrawal review of about 3 business days, and Proximal invoices monthly with undisputed amounts due within 30 days. Trigger-by-trigger terms are set out under when and how AI data buyers pay.

Selling an Environment: Benefits, Risks, Downsides and Alternatives, Including Not Selling

Selling an environment turns a reusable build into revenue, but the license usually survives any later change of mind and the buyer side is thin.

Benefits

• Demand: one startup employee told Epoch AI (2026-01-12) of 10 to 20 times more RL requests than benchmarking requests, a claim of that interviewee.
• Ownership: DataVendor states the usual starting point is that the vendor keeps ownership, which leaves further non-exclusive sales possible (checked 2026-11-09).
• Exits: Mercor.io Corporation acquired Sepal AI, Inc. (2026-02-06), so acquirers take on whole builders.

Risks

• Weak graders get gamed, as METR's 30.4% o3 figure (2025-06-05) shows.
• A buyer can pull back: Anthropic's S-1 (2026-10-26) states it suspended higher-risk RL environments after 2026 incidents in pre-release cybersecurity evaluations.
• Counterparties change through acquisitions, and Mercor's Deeptune announcement (2026-07-09) has no confirmed closing as of 2026-11-09.

Downsides

• Sales are final, and the buyer's entitlement survives unpublishing (DataVendor, checked 2026-11-09).
• DataVendor takes a 20% fee at withdrawal and sets a USD 1,000 minimum withdrawal (checked 2026-11-09).
• Open hubs state no payment to creators (OpenReward, BenchFlow).

Alternatives

• License the underlying data or software to a builder that turns it into environments: Abundant, Halluminate and Plato each invite company software or data, and none publishes payment or license terms (checked 2026-11-09).
• Publish the environment open source on a hub such as OpenReward.
• Keep the environment for internal evaluation.
• Not selling at all is a valid outcome, weighed in should you sell or license your data for every data type.

Where Selling an Environment Meets Builders, Code, Trajectories and the Seller Guide

An environment sale borders three neighbors: the builders who buy and resell environments, the codebases and trajectories sold as data on their own, and the general seller guide.

• Selling or licensing source code and codebases as data, for the code that seeds tasks.
• Computer-use and agent trajectory data sold on its own.
• The complete guide to sell or license company data to AI companies.
• Choosing a buyer for your data among labs, vendors and marketplaces.
• AI training data marketplaces, fees compared.
• Human data for AI: experts who write tasks and rubrics.
• How DataPacts is funded: no fees, no cut of any sale.
• What RL environments sell for: dated public price points.
• RL environments for coding agents.

Frequently Asked Questions

Can an individual get paid for writing RL tasks?

Yes, through task bounties, which pay per accepted task rather than for an environment. Refresh states USD 30 to 75 per task and Shipd shows USD 80 to 220 per item (checked 2026-11-09). Pay is compared under AI training jobs and platforms compared.

Do AI labs buy RL environments from small builders?

Labs do not publish their purchases from named builders. Among labs' own filings, only Anthropic's S-1 (2026-10-26) speaks of RL environments, and lab names elsewhere are builders' own claims.

Can the same environment be sold to more than one lab?

Yes if the license is non-exclusive and the seller keeps ownership. DataVendor states that a sale licenses rather than transfers, while Proximal's agreement lets the builder reuse deliverables (checked 2026-11-09).

How much does an RL task sell for?

No price list exists; stated figures are claims. Epoch AI's interviewees state USD 200 to 2,000 per task (2026-01-12) and Mechanize about USD 500 (2025-08-18). Other data types are priced under AI training data prices for other data types.