RL Environments: What They Are, Who Builds Them and Who Buys Them

RL environments are simulated workplaces, software replicas and task sets in which AI agents practice tasks and are scored, and AI labs buy them as training data. Anthropic's S-1 of 2026-10-26 names datasets licensed or commissioned from third parties among its training data sources and states that it suspended its higher-risk RL environments after 2026 incidents in pre-release cybersecurity evaluations. This hub lists the 31 builders DataPacts profiles, alphabetically, with the legal entity on record and the date each row was checked.

Within AI training data licensing, RL environments are a product that changes hands like a dataset. Mercor.io Corporation acquired the builder Sepal AI on 2026-02-06 and announced on 2026-07-09 that it will acquire Deeptune, with the closing not confirmed as of 2026-11-09. Public price claims from interviewees and builders disagree, and builders' published terms differ most on who keeps the finished environment.

What is an RL environment in AI training?

An RL (reinforcement learning) environment combines a task, a starting state, tools to act with and a verifier that scores the result, and the AI agent placed in it learns from those rewards rather than from fixed examples. The verifier, also called a grader or rubric, turns each outcome into a reward signal, and the recorded steps of each run form a trajectory. Domain experts write many of the tasks and rubrics, which is why companies that supply human experts also build and sell environments.

The table breaks an environment into its five components, with an example from a builder's own page for each, as of 2026-11-09.

Difficulty decides whether an environment teaches anything. Interviewees quoted by Epoch AI on 2026-01-12 put the minimum useful pass rate at about 2 to 3 %, or one success in 64 to 128 rollouts, and Chakra Labs wrote on 2026-01-29 that labs expect pass@5 below 30 %. Tasks that are too easy give no signal, and tasks that are never solved give none either.

Environments also fail when the model games the scoring check instead of solving the task. METR measured reward hacking by OpenAI's o3 in 30.4 % of RE-Bench runs (39 of 128), published on 2025-06-05, which is why builders sell the quality of the grader as much as the task itself.

Environments, tasks and trajectories as data acquired for AI training

In the training-data market an RL environment is a product: a lab licenses or commissions the environment, its task set and the trajectories agents leave in it, much as it licenses a dataset. The three parts can be sold together or apart, and each carries its own rights question.

Anthropic's S-1 of 2026-10-26 lists its training data sources as public web data, datasets licensed or commissioned from third parties, and synthetic data generated by its own models, which places commissioned environments beside licensed corpora. Trajectories also trade on their own: micro1 states that it purchases agent trajectories for USD 1 to 2 million annually, checked 2026-11-09.

Builders in turn acquire data to make environments realistic. Abundant launched Foundry with Fireworks AI on 2026-11-03 to build custom environments from a customer's user sessions, tools and golden test cases, and offers a "License your data" contact option. Good Start Labs helps game publishers license gameplay data to AI labs, Halluminate and Plato invite companies to offer data and workflows, and DataVendor lists private codebases beside tasksets (all checked 2026-11-09). The inputs on record are codebases with Git history, company records, gameplay data, longitudinal clinical records and user sessions.

The seven types of RL environments sold to labs

Builders sell seven recurring types of RL environments: coding, computer use, enterprise workflow replicas, games, healthcare, safety and red-teaming, and open environment hubs. The grouping follows what each builder states on its own pages, and one builder can sit in several rows, as Huzzle Labs does with coding, computer-use and enterprise-workflow environments.

Gray Swan AI sits under safety because it is an AI security company that runs adversarial evaluations and a crowdsourced red-teaming Arena, not a general RL-environment vendor. Open hubs differ from the other six types: they host environments contributed by others rather than build them to order. Prime Intellect launched its Environments Hub on 2025-08-27 and states that more than 400 environments were crowdsourced in about two months (2025-10-27), and General Reasoning states that OpenReward carried more than 330 environments at its launch in March 2026. Halluminate, by contrast, builds to a domain: its Westworld Finance Diligence Bench holds 88 problems, as Halluminate stated on 2026-08-11.

The table assigns builders from the directory below to each type, as of 2026-11-09, with the directory rows as the source.

Who builds RL environments: the builder directory

DataPacts profiles 31 RL environment builders, listed alphabetically with the legal entity on record, what each builds or sells and the date its primary pages were checked. Three terms in the table need a key.

• Not published: the builder names no legal entity on any primary page checked on 2026-11-09.
• Identity not confirmed: five rows (Arga Labs, Idler, Matrices, Preference Model and Vmax) carry no attribute and no profile link until the company's legal identity is confirmed.
• Mercor transactions: Sepal AI was acquired by Mercor.io Corporation on 2026-02-06, while the Deeptune acquisition was announced on 2026-07-09 and its closing is not confirmed as of 2026-11-09.

The directory below lists every builder with its source domain, as of 2026-11-09.

Registry numbers come from SEC EDGAR Form D filings and Companies House: six builders carry an SEC CIK and two a Companies House number. Two entity notes matter for contracts. Datacurve's Contributor Terms name Datacurve AI, Inc. while its privacy policy names Datacurve, Inc., and Plato Technologies, Inc. is not the same issuer as the EDGAR filer "Plato Technologies Inc" of Wyoming and New York.

AIChamp stays in the directory with its current business stated: it formerly marketed RL environments as "Virtual Gyms" and since 2026 sells factory loss studies and digital twins under terms effective 2026-10-22.

Some builders also want company records as inputs and act as AI training data buyers in their own right. Builders that list the records they acquire, such as Collinear AI, Ooak Data and pre.dev, are profiled among AI training data buyers rather than in this directory.

Who buys RL environments

Frontier labs are the end buyers of RL environments; human-data companies buy builders outright, and marketplaces sell tasksets to invited labs.

Frontier labs: what filings and builders' pages show

The only frontier-lab filing on record that addresses RL environments is Anthropic's S-1 of 2026-10-26, which states that Anthropic suspended its higher-risk RL environments after 2026 incidents in pre-release cybersecurity evaluations. Every other lab name comes from a builder's own page.

The table lists each buyer named or counted on record, ordered by date, as of 2026-11-09. Lab names other than Anthropic's own filing are each builder's claim.

Gray Swan's figures describe a security business: its Arena counts more than 15,000 participants. Two of the claims name no lab at all: Halluminate counts four of the five leading closed-source U.S. labs without naming them, and BioStack counts "top AI labs" among its 17 customers and prospects. Andon Labs' "Trusted by" logos are a display, not a contract, and Project Vend, its research partnership with Anthropic, states no data license or training-data transfer, so it is not a purchase. No lab has stated an RL-environment purchase from a named builder in a filing or on its own pages on record as of 2026-11-09; Anthropic's S-1 speaks only of its own RL environments and of commissioned datasets.

Human-data companies that acquire builders or build environments

Mercor.io Corporation acquired Sepal AI on 2026-02-06 and announced on 2026-07-09 that it will acquire Deeptune; the Deeptune closing is not confirmed as of 2026-11-09. Orrick, which advised Mercor, announced the Sepal AI deal with the price undisclosed, and sepalai.com returns 404. Mercor's announcement states that the Deeptune team joins Mercor in New York and gives no value, while Deeptune's site says that it is joining Mercor. The company data program behind both moves is described in the profile of Mercor as a data buyer.

Other human-data companies sell RL-environment work or source its inputs, among them:

• micro1: states purchases of agent trajectories at the annual range given above, checked 2026-11-09.
• Scale AI: states that its RL Environments offering holds 2,000+ tasks, four enterprise environments and 370 OSWorld Advanced tasks, checked 2026-11-09.
• Turing: Project Lazarus pays holders of company records a revenue share with nothing upfront, and Turing states that the holder also earns when Turing builds RL environments on the record; the split is not published (checked 2026-11-09).

Marketplaces where labs buy tasksets and environments

DataVendor, run by HUD, lists codebases, raw data, HUD tasksets and RL environments that vendors price themselves, for buyers invited from a waitlist. A sale licenses the assets for a commercial use rather than transferring them, and the vendor usually keeps ownership.

HUD's own site counters showed 13,764 environments built and 4,831,034 task runs when checked on 2026-11-09. Only organizations may sell, and they sign up without an application review. DataVendor states that it keeps 20 % of earnings, charged at withdrawal and never on referral fees, with a minimum withdrawal of USD 1,000, checked 2026-11-09. The full seller terms sit in the DataVendor (HUD) profile.

Price claims and contract structures reported for RL environment work

No builder publishes a price list for RL environments; the public figures are statements by interviewees, builders and marketplaces, and they disagree. None of them is a market price, and none is averaged here.

The table sets each public figure beside the party that stated it, as of 2026-11-09.

The figures disagree at the low end. Interviewees quoted by Epoch AI on 2026-01-12 put a task at USD 200 to 2,000, while Mechanize wrote on 2025-08-18 that about USD 500 per RL task is the minimum efficient spend, because compute alone costs about USD 2,400 per task over five reuses. A task priced at the bottom of the Epoch range therefore costs less than Mechanize's stated floor, and Mechanize argues that such cheap tasks waste compute.

Contract structure varies more than unit price. Epoch AI's interviewees described contracts of six to seven figures per quarter, two founders put exclusivity at roughly 4 to 5 times the non-exclusive price, and one startup employee reported 10 to 20 times more RL requests than benchmarking requests. DataVendor states that some single-codebase deals exceeded USD 10 million, checked 2026-11-09. Refresh states that the price of a contributed task is agreed upfront and settled on approval, with larger scopes quoted separately.

Dated price points by environment type are gathered under what RL environments sell for.

Who owns what: builder terms on deliverables, reuse and contributor work

Builders' published terms differ most on who keeps the environment: Proximal keeps its deliverables, DataVendor sellers keep ownership, and Shipd contributors assign their work outright. The pattern repeats across hosting hubs and bounty programs.

The table compares seven published documents, as of 2026-11-09, with each builder's own page as the source.

The contrast matters to a data holder. Under Proximal's MSA the customer holds only an internal-use license during the Term, while AIChamp's terms effective 2026-10-22 keep customer data inside the project and delete it within 30 days. Contributor platforms differ again: Shipd contributors assign their work on creation or submission, while General Reasoning states that environment code stays on GitHub. Prime Intellect states that its entire stack is open-source, and its verifiers library carries the MIT license, so environments built on it stay reusable by anyone. The clauses that set these outcomes are explained under AI data licensing agreements.

Benefits, risks, downsides and alternatives for data holders, including not licensing at all

A company whose records, code or software could seed an RL environment can license them to a builder, but the license usually outlives the project and the model trained on it.

Benefits:

• Revenue from dormant records: Halluminate, Plato and Abundant run company data programs, and Turing states a revenue share when it builds environments on a record (checked 2026-11-09).
• Environments from the company's own workflows: Abundant's Foundry starts from a customer's user sessions, tools and test cases.
• Contained use: AIChamp's current terms keep customer data inside the project and rule out selling or licensing it.

Risks:

• Reuse by the builder: Proximal's MSA s.8.1 lets Proximal reuse Deliverables for other customers when they contain no customer confidential information.
• Personal data: DataVendor cleans purchased repositories and raw data of personal information within 2 business days, but not tasksets and environments.
• A new counterparty: acquisitions change who holds the environment and its inputs, as with Sepal AI and the announced Deeptune deal.

Downsides:

• Unpublished terms: Halluminate, Plato and Good Start Labs publish no license or payment terms, checked 2026-11-09.
• No way back: DataVendor states that sales are final, with no refunds, and that a buyer's entitlement survives unpublishing.

Alternatives:

• License records to a company data partnership program that pays for operating records rather than environments.
• List a codebase on a marketplace such as DataVendor and keep ownership under a license.
• Contribute to open hubs, BenchFlow, General Reasoning or Prime Intellect, where only Prime Intellect states bounties for contributions.
• Do not license at all: the records stay out of every environment and every model, and the guide to sell or license company data to AI companies treats declining as a valid outcome.

How this hub is built: neutral order, check dates and no fees

The builder directory is alphabetical and includes every builder DataPacts profiles, whether or not it pays referral fees, and 26 of its 31 rows carry the date their primary pages were checked. The other five show n/a because the company's legal identity is not confirmed. Competitor directories of RL-environment companies are never a source, and referral programs under which no data changes hands, such as AIChamp's referral for manufacturing clients, are left out.

Every figure on this page names the party that stated it and the date, following how we source, date and check every figure. Funding rounds reported only by the press or by investors are not stated for any builder.

Where RL environments meet human data, synthetic data and the seller path

RL environments sit between human data, which supplies experts and tasks, and synthetic data, which models generate for themselves. Related pages cover each neighbor and each step of the seller path.

• Human data for AI: experts who write tasks and rubrics, and their pay.
• Company data partnership programs: programs that license operating records.
• Program terms compared: payouts and program terms, dated.
• AI data licensing deals: Mercor's acquisitions among the dated deals.
• How DataPacts is funded: no fees, no cut of any sale.
• RL environment companies: builders and vendors.
• Sell an RL environment or task set: the steps for builders.
• Rights in an RL environment: cloned software, seed data and contributor IP.
• Synthetic data for AI training: what it is and its limits.
• RLHF data: what reinforcement learning from human feedback needs.

Frequently asked questions

What does RL environment mean?

An RL environment is a simulated task space, with tools and a scoring check, in which an AI model learns by reinforcement learning. Interviewees quoted by Epoch AI on 2026-01-12 put the minimum useful pass rate at about 2 to 3 %, so tasks must be hard.

Which startups build RL environments?

DataPacts profiles 31 builders, listed alphabetically in the directory above, from Abundant to Vmax. They include coding specialists such as Mechanize and Proximal, computer-use builders such as Plato, enterprise replica builders such as Deeptune and Fleet AI, and open hubs such as Prime Intellect.

Did Mercor buy RL environment companies?

Mercor acquired one builder and announced a second acquisition. Mercor.io Corporation acquired Sepal AI on 2026-02-06, with the price undisclosed. On 2026-07-09 Mercor announced that it will acquire Deeptune, and the closing is not confirmed as of 2026-11-09.

How much do RL environments cost?

No price list exists, and the public claims disagree. Epoch AI interviewees put tasks at USD 200 to 2,000 (2026-01-12), while Mechanize states about USD 500 per task as the minimum efficient spend (2025-08-18). Other data types are compared under AI training data prices.

Can a company license its own data to an RL environment builder?

Yes, several builders invite company data, though none of their contact-based programs publishes payment terms. Halluminate, Plato and Abundant run contact-based programs, checked 2026-11-09, and DataVendor lets organizations list codebases. Methods for pricing records are explained under what company data is worth to AI companies.

Is an RL environment the same as a benchmark?

No: a benchmark measures a model, while an RL environment trains it with rewards. The same tasks can serve both, which is why Andon Labs and Gray Swan AI build evaluations, and one startup employee told Epoch AI of 10 to 20 times more RL requests. How labs buy AI evaluation data is covered separately.