RL Environment Companies: Builders and Vendors

RL environment companies build and sell the simulated workplaces, software replicas and task sets that AI labs use to train agents, and DataPacts tracks 31 of them by business model, legal entity and dated record. Eight of the 31 appear in a public registry by number, ten have a funding figure stated by the company or a filing, and Mercor.io Corporation acquired 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. Anthropic's S-1 of 2026-10-26 lists datasets licensed or commissioned from third parties among its training data sources.

This directory sits within AI training data licensing, is alphabetical and carries no ranking. Every row was read from the company's own page, a filing or a registry, and how we source, date and check every figure is set out on the methodology page.

What Counts as an RL Environment Company

A company counts as an RL environment company here when it builds, hosts or sells environments, task sets or trajectories that AI developers use for training, not only compute or test scores. The test is what the company's own records say it sells, not how it describes its market. A company that appears in a registry but sells only compute, or one that publishes benchmark scores without offering training material, does not meet it. The test is applied row by row, and a company whose records change is moved between segments or dropped when the change is dated.

• Counted: companies that build or sell environments, task sets or trajectories for training.
• Not counted: sandbox infrastructure providers, which supply compute for others' environments, and evaluation-only firms.
• Not counted: research environments published without a commercial offer.

The parent page covers what RL environments are and which labs buy them, so this page does not repeat the definition or the buyer evidence. Two evaluation companies stay on the list only because their own records show training data, and the table of record conflicts below states why.

The RL Environment Company Landscape by Business Model

The 31 companies fall into seven segments by how they make money: custom builders for labs, builders working from a holder's data, platforms and hubs, task contribution platforms, evaluation companies, acquired or repositioned companies, and companies whose identity is not confirmed. The table maps each segment to its members as of 2026-11-09.

Segments describe the main business line of each company and a company sits in one segment only. Datacurve sits with the custom builders although its Shipd site pays contributors in the same way as Refresh. The five companies whose identity is not confirmed are not placed by business model, because no attribute is stated for them. Where a company runs two lines, such as a lab-facing service and a contributor platform, the segment follows the line its own pages describe first.

All 31 RL Environment Builders and Vendors, With Legal Entity

The table lists every RL environment company DataPacts tracks, alphabetically, with the legal entity and registry on record, the year and base the records give, and its status, checked November 9, 2026. Each linked name leads to the company's own profile once that profile is published. The year and base are the ones the records give: a Form D states the state and year of incorporation, a Y Combinator page states the batch, and a footer states an office. A field with no such source reads "not stated" and is never filled from press coverage.

• Not published: the company's pages show no legal name.
• Identity not confirmed: the name is listed with no attribute until the company is verified.
• Not stated: no primary source gives that field.

Legal Entities, Registries and Where Company Records Disagree

Eight of the 31 companies appear in a public registry by number, six as SEC Form D issuers and two at Companies House, and for the rest the legal name comes from the company's own terms or footer. The six Form D issuers are Bespoke Labs, Chakra Labs, Deeptune, General Reasoning, Gray Swan AI and Prime Intellect; the two Companies House entries are Eigent AI and Huzzle Labs.

Where two records for one company differ, the page shows both and does not pick one. A Form D can cover one closing of a larger round, so a difference in amount is not an error by itself, and a company post and a filing can both be accurate. Neither is corrected to match the other, and a reader who relies on one record for a contract should check both. The table lists the six cases, as of 2026-11-09.

Funding Stated by the Companies or Their Filings

Ten of the 31 companies have a funding figure stated by the company itself or recorded in an SEC filing, and each is written here as that party's statement with its date. The table gives the stated figure, the stating party and any revenue or scale claim, as of 2026-11-09.

Revenue and scale figures are the company's own claims and no filing confirms them. Funding amounts reported only by the press or by an investor are not stated on this page, and a company missing from the table has no figure from the company or a filing. A funding round says what investors paid for shares, not what any lab pays for an environment, so none of these figures is a price. Prices claimed for RL environment work are outside this page.

Acquisitions, Exits and Repositioning Among RL Environment Companies

Two RL environment companies have moved to Mercor in 2026: Mercor.io Corporation acquired Sepal AI on February 6, 2026, and announced on July 9, 2026 that it will acquire Deeptune, a deal whose closing is not confirmed as of November 9, 2026. The table dates each change from the stating party, and "announced" is kept apart from "acquired" throughout. No party has announced an acquisition of Mechanize, so none is listed. The Sepal AI website returning HTTP 404 on 2026-11-09 is a DataPacts check, not a statement by Mercor or Sepal AI.

Both Mercor transactions are acquisitions of builders, and neither record describes a Mercor environment product. For the buyer side of the same company, see Mercor's profile as a data buyer.

Human-Data Companies and Data Buyers That Also Build RL Environments

Some companies that build RL environments are profiled as data buyers, because they acquire company records, codebases or app history to build them. Three human-data companies describe environment work in their own program terms, each as that company's claim, checked 2026-11-09:

• Scale AI: its RL Environments page states 2,000+ tasks and 4 enterprise environments.
• micro1: states, checked 2026-11-09, that it licenses qualifying agent trajectories for USD 1 to 2 million annually.
• Turing: Project Lazarus pays a revenue share with nothing upfront, and the holder is also paid when Turing builds RL environments on the record; the split is not published.

The table lists the five builders with their own buyer profile, alphabetically, as of 2026-11-09.

Every buyer and program on the site, built the same way, is compared in AI training data buyers compared.

Checking an RL Environment Company Before Licensing Data or Contributing Work

A data holder or contributor can check an RL environment company in four steps, starting with whether the legal entity named in its contract exists in a registry.

1. Look up the entity in a registry: SEC EDGAR for a Form D, Companies House for a UK company.
2. Compare the legal name across the terms, the privacy policy and the contract, because the names can differ, as with Datacurve.
3. Read the published terms on deliverables and assignment: Proximal's MSA states that Proximal owns deliverables and may reuse them (section 8.1), and Shipd's contributor terms assign all work product to Shipd, which Datacurve AI, Inc. operates (section 3.1).
4. Set the company's claims against its filings, as in the Prime Intellect and Deeptune rows above.

The first step also settles the counterparty, and choosing a buyer for the data covers the rest of the selection.

Benefits

• Programs run by Abundant, Halluminate, Plato and Good Start Labs invite a holder's data or software, which can turn existing records into a licensed asset.
• Open hubs such as General Reasoning's OpenReward and BenchFlow's PostTrain Arena state no assignment of the contributor's work.

Risks

• An acquisition changes the counterparty: the Sepal AI and Deeptune deals moved or will move builders under one acquirer.
• A builder may keep the deliverables, as Proximal's MSA states.
• Entity names can differ between documents of one company, so the contract may name a different entity than the website.

Downsides

• Abundant, Halluminate, Plato and Good Start Labs do not publish the terms of their programs.
• Open hubs state no payment to environment creators.

Alternatives

• Licensing through a company data program described in plain text by the buyer.
• Listing the data on a marketplace, where the vendor sets the price: DataVendor (HUD) charges a 20 % fee on earnings at withdrawal, as HUD's documentation states, checked 2026-11-09.
• Not licensing or contributing at all, which remains a valid outcome.

How This Company List Is Built: Inclusion, Order and Holds

The list is alphabetical, includes every company whose records show RL environment work for AI training, and never depends on referral fees. It has no score, no ranking by revenue and no paid placement, and competitor directories are never a source of a fact.

Five names carry no attribute because their identity is not confirmed, and three companies show "not published" because their pages give no legal name. Those rows are updated when a primary page names the entity. The method behind every date and figure is on the methodology page.

Where the Company List Meets Environment Types, Selling and Data Types

This page lists the companies; what the environments are, how to sell one and which data types feed them sit on their own pages.

• Environment types, buyers and terms: RL environment types, buyers and terms, with RL environments for coding agents, RL environments for computer-use agents, and enterprise workflow RL environments. How an environment is scored is covered under verifiers, graders and rubrics in RL environments.
• Selling and rights: how to sell an RL environment or task set to AI labs, public price points for RL environments and cloned software, seed data and contributor IP.
• Data types: computer-use and agent trajectory data, licensing source code and codebases to AI companies and human data companies and expert contributors.
• Terms and deals: payouts and program terms compared, dated, the AI data licensing deal tracker and DataVendor, the marketplace HUD runs.

Frequently Asked Questions

Which are the top RL environment companies?

DataPacts does not rank RL environment companies; the list is alphabetical, and only three publish revenue figures, each as their own claim. Halluminate, Prime Intellect and Proximal state revenue or run-rate figures on their own pages, and no filing confirms them.

Do RL environment companies buy company data?

Some do: Abundant, Halluminate, Plato and Good Start Labs run programs for a holder's data or software, and five more builders are profiled among data buyers.

The programs and their published terms are compared under company data partnership programs.

Are evaluation and sandbox companies RL environment companies?

Sandbox infrastructure is not counted, and evaluation companies count only where their records show training data.

Andon Labs and Gray Swan AI are listed on that basis, and the evaluation page explains how labs buy evaluation data.

Do RL environment companies hire contributors?

Several pay contributors per task, among them Refresh, Datacurve through Shipd and Prime Intellect through bounties, each at amounts it states itself.

Jobs and pay across the sector are compared under AI training jobs and platforms compared.