RL Environment Companies
noun · AI trainingA company that builds reinforcement learning environments: sandboxes where an AI agent attempts tasks with real tools while a grader scores each attempt. Labs train agents on these scores, so the environment decides what the agent learns.
Companies (14)
- Bespoke Labs builds RL environments and the engine that runs them.
- Halluminate managed sandboxes modeled on Salesforce, Slack and ticketing tools.
- Patronus AI simulated copies of websites and in-house systems, its "digital world models".
- Scale AI tool-use and computer-use environments with published task counts.
- Turing Jira, Salesforce and Zendesk interface copies, and MCP backends.
- Surge AI RL environments, rubrics and verifiers from a human data company.
- Mercor expert network with RL environments; acquiring Deeptune.
- micro1 managed human experts producing post-training and RL data.
- Handshake AI students and experts on paid AI lab projects; named by Epoch AI.
- Toloka auto-verifiable tasks for RL and evaluation, from Amsterdam.
- Appen ready-made coding tasks, each with an automatic check.
- Centific RL environments inside its AI Data Foundry platform and services.
- Prime Intellect runs the open Environments Hub for sharing RL environments.
- OpenPipe RL training tools and the open-source ART library; part of CoreWeave.
A listed company can pay us a referral fee if one of our introductions turns into a signed deal. Inclusion and wording do not depend on it.
What does an RL environment contain?
Four parts, sold together. All four sit inside a sandbox that resets to the same start for every attempt.
Tasks
A starting state and an instruction, such as fixing a failing test or closing a support ticket. Training needs many, varied enough that the agent learns the skill instead of the answers.
Tools
What the agent may do: run a shell, click through a browser, call an app's API. The tool set is part of the product and should come documented.
A grader or verifier
Code, a rubric or a reviewer that checks the end state and says pass or fail. It is the hardest part to build well, because agents exploit any gap it leaves.
Rewards
The score the training loop reads, usually taken from the grader: a plain pass or fail, or partial credit for steps on the way.
The full breakdown, with what to ask a vendor about each part, is on what is in an environment.
Who builds RL environments?
Three kinds of company. Specialists build environments and little else. Large human data providers have added environments to their RLHF and evaluation work. A third group runs open hubs and training tools rather than building to order.
| Company | Kind | What it says it builds | Published price |
|---|---|---|---|
| Bespoke Labs | Environment specialists | RL environments plus the infrastructure to run them: an environment engine, sandboxing and execution, agent optimization. Sells to frontier labs, newer labs and enterprises. | Not published |
| Halluminate | Environment specialists | Hosted practice copies of business software (its examples: Salesforce, Slack, ticketing systems) where agents that drive a computer or browser are trained and tested. | Not published |
| Patronus AI | Environment specialists | Simulated copies of websites and company systems, which it calls digital world models, used to train agents and to stress-test them; also automated agent evaluation. | Not published |
| Scale AI | Data providers that also build environments | Enterprise tool-use environments and computer-use environments, alongside its wider data and evaluation work. | Not published |
| Turing | Data providers that also build environments | Copies of the Jira, Salesforce and Zendesk interfaces for agents to work in, plus backend environments reached over MCP, within its post-training and evaluation business. | Not published |
| Surge AI | Data providers that also build environments | RL environments, rubrics and verifiers, alongside RLHF, demonstrations and human evaluation for AI labs. | Not published |
| Mercor | Data providers that also build environments | Expert-built training data, human evaluation, the APEX benchmarks and RL environments for AI labs. | Not published |
| micro1 | Data providers that also build environments | Post-training and RL data produced by human experts it recruits and manages. | Not published |
| Handshake AI | Data providers that also build environments | Its fellowship pays students and subject experts, recruited through the Handshake jobs network, to work on projects that train and evaluate AI lab models. | Not published |
| Toloka | Data providers that also build environments | Auto-verifiable tasks for evaluation and RL, plus demonstrations, preference data and red teaming. | Not published |
| Appen | Data providers that also build environments | Ready-made RL tasks that come with automatic checks, sold beside its annotation and post-training data. | Not published |
| Centific | Data providers that also build environments | RL environments as one part of its AI Data Foundry platform and services, with preference data, fine-tuning data and safety evaluation. | Not published |
| Prime Intellect | Open hubs and training tools | The Environments Hub, where anyone can publish or reuse RL and evaluation environments; also the prime-rl training code (open source) and a paid hosted trainer. | Not published |
| OpenPipe | Open hubs and training tools | Software that uses reinforcement learning to train agents, and the open-source ART library. Tools rather than built-to-order environments. | Not published |
Referral fees from any of these companies, where an introduction ends in a contract, play no part in the order or the descriptions.
Sources, task counts and dated changes for each company are on the companies page.
What does one RL task cost to build?
That is the range Epoch AI heard from people in the market, for one task together with its grader. It put $20,000 for a single task as rare but possible. Epoch's figures come from interviews, not contracts, and no vendor publishes a rate card.
Price moves with how hard a task is to grade, how much expert time goes into writing it, and whether you want the rights to yourself. The environment budget worksheet turns a task count into a range on this basis and lists what to ask; task costs has the detail.
Exclusive or shared rights?
An exclusive environment is built for one buyer and not resold. A shared (non-exclusive) one may be licensed to other labs too. Keeping a build to yourself carries a premium; in Epoch AI's interviews it came to 4 to 5 times the shared price.
Shared rights suit common skills, such as general coding or browsing, where the edge comes from how you train rather than what you train on. Exclusive rights make sense when the tasks encode something particular to your product. Either way, put in writing who owns the task specs, the graders and the trajectories your runs produce.
Where do the workflows come from?
Business-tool environments copy how real work is done: the screens of a CRM, the steps of closing a ticket, the checks in an invoice run. Builders get this from domain experts who describe and perform the work, from replicas of public software, and sometimes from companies that license records of their own operations after removing customer details. The closer a task follows real practice, the more of what the agent learns carries over to the job itself.
Where these facts come from (15)
- Epoch AI, 12 January 2026: "An FAQ on reinforcement learning environments" (Denain and Barber) (page checked 7 November 2026).
- Bespoke Labs blog, funding announcement (6 Jul 2026) (page checked 7 November 2026).
- Y Combinator launch page for Halluminate (page checked 7 November 2026).
- Patronus AI website (page checked 7 November 2026).
- Scale AI RL environments page (page checked 7 November 2026).
- Turing RL environments page (page checked 7 November 2026).
- Surge AI products page (page checked 7 November 2026).
- Mercor blog, Mercor to acquire Deeptune (9 Jul 2026) (page checked 7 November 2026).
- micro1 website (page checked 7 November 2026).
- Epoch AI's January 2026 RL environments FAQ, which names it as a seller (page checked 7 November 2026).
- Toloka about page (page checked 7 November 2026).
- Appen blog on RL with verifiable rewards (page checked 7 November 2026).
- Centific AI Data Foundry page (page checked 7 November 2026).
- Prime Intellect blog, Environments Hub launch (27 Aug 2025) (page checked 7 November 2026).
- CoreWeave news, acquisition of OpenPipe (3 Sept 2025) (page checked 7 November 2026).
Questions about RL environment companies
- What do RL environment companies sell?
- Sets of tasks inside a resettable sandbox, the tools an agent may use there, and graders that score each attempt. Some deliver code you run yourself; others host the environment.
- Do any of them publish prices?
- None of the fourteen companies listed here publishes a price for environments or tasks (checked 7 November 2026). Every one works on custom quotes.
- What does one RL task cost?
- In most cases $200 to $2,000 for a task and its grader, going by Epoch AI's interviews, with $20,000 rare. These are estimates, not price lists. See task costs.
- Is an exclusive environment worth the premium?
- Expect to pay several times more: Epoch AI heard a multiple of 4 to 5. Exclusivity matters most when the environment encodes a skill your competitors would also want.
- How is an RL environment different from a dataset?
- A dataset is fixed examples. An environment is interactive: the agent acts, the world changes, and the grader scores the outcome, so the same environment yields new training signal on every run.
- Can a small team buy an RL environment?
- Yes. Several vendors sell to enterprises and newer labs as well as frontier labs, and open hubs offer free environments to start from. Starting with a small pilot set lets you check grader quality before you commit to a full order.
Get a quote
Tell us the company ZIP code, a contact, and an outline of the environments you have in mind. We compare it with what each builder publishes and send you the shortlist by name.
How the site earns: we make introductions, and a builder you later contract with may pay us for yours.
Received. The next three things:
- Your outline is set against what each RL environment builder says it makes and the terms it publishes.
- A shortlist arrives by email: names, links, and a sentence on the fit for each.
- No builder hears about you until you answer yes. There is no mailing list to join.