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 and not the answers.
Fourteen builders · buyer side
RL Environment Companies
The companies that build sandboxes, tasks and graders for training AI agents. Tell us what your agent has to learn, and we introduce the builders whose work fits it.
Task
A start state and an instruction
task_0193 helpdesk replica start: 3 open tickets ask: refund order 4471, then close T-118Agent run
The agent acts with real tools
open_ticket("T-118") orders.lookup(4471) refund(4471, 42.00) set_status("T-118", "closed")Grader
Code checks the end state
- Refund row in orders table
- T-118 status is closed
- Reply sent to customer
Reward
The score training reads
0.67
Partial credit: two of three checks passed. A plain pass or fail would read 0.
Invented example task and run. The ticket, order and amount are made up.
Four parts sold together, inside one sandbox
The sandbox resets to the same start for every attempt. The full breakdown, with what to ask about each part, is on what is in an environment.
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.
- Editing the failing test instead of the code
- Creating a new record instead of updating one
- Padding answers for a model judge
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.
Where the workflows come from
Business-tool environments copy how real work is done: a CRM's screens, the steps of closing a ticket, the checks in an invoice run. Builders learn them from domain experts, from replicas of public software, and sometimes from companies that license records of their own operations with customer details removed.
The few numbers anyone has published
None of the fourteen companies prints a price for environments or tasks. What is public comes from Epoch AI's interviews with people in the market, and from task counts on vendors' own pages.
- $200–$2,000One task with its grader, in most cases (Epoch AI)
- $20,000A single task, rare but possible (Epoch AI)
- 4–5×Exclusive rights against shared (Epoch AI)
- 2,000+Tasks in Scale AI's Enterprise Atlas environments
- 500Coding tasks in Appen's set, each with a verifier
As Epoch AI and each company publish them, from interviews and product pages rather than contracts. Not offers.
Sources for every figure and company (15)
- Epoch AI, 12 January 2026: "An FAQ on reinforcement learning environments" (Denain and Barber).
- Bespoke Labs blog, funding announcement (6 Jul 2026).
- Y Combinator launch page for Halluminate.
- Patronus AI website.
- Scale AI RL environments page.
- Turing RL environments page.
- Surge AI products page.
- Mercor blog, Mercor to acquire Deeptune (9 Jul 2026).
- micro1 website.
- Epoch AI's January 2026 RL environments FAQ, which names it as a seller.
- Toloka about page.
- Appen blog on RL with verifiable rewards.
- Centific AI Data Foundry page.
- Prime Intellect blog, Environments Hub launch (27 Aug 2025).
- CoreWeave news, acquisition of OpenPipe (3 Sept 2025).
Builders
Three kinds of company build them
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 instead of building to order.
Mercor is buying Deeptune, an environment specialist, so we file Deeptune under Mercor. Sources and dated changes for each company sit on the companies page.
Environment specialists
Data providers that also build environments
Open hubs and training tools
A listed company can pay us a referral fee if one of our introductions turns into a signed deal. Inclusion, grouping and wording do not depend on it.
Cost
Budget per task, not per environment
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. Epoch's figures come from interviews, and no vendor publishes a rate card.
The environment budget worksheet turns a task count into a range on this basis; task costs has the detail.
Per task, with its grader
- Most tasks
- $200 to $2,000
- Rare
- $20,000
Log scale. Epoch AI interview figures, not quotes.
Rights
Exclusive costs several times more. Sometimes it is worth it
An exclusive environment is built for one buyer and not resold; a shared one may be licensed to other labs too. Shared rights suit common skills such as general coding or browsing, where the edge comes from how you train. Exclusive rights make sense when the tasks encode something particular to your product.
Put these in writing either way
- Who owns the task specs
- Who owns the graders, and who fixes them
- Who owns the trajectories your runs produce
- Whether the build may be resold, and when
Test the grader before you sign
The grader is the part of an environment to examine most closely. Five checks to run on any vendor's pilot set.
- 1
Read a sample
Tasks with their graders, in a form your own engineers can read line by line.
- 2
See what the grader rejected
The deliberately wrong solutions each grader was tested against, and the share of tasks that failed review.
- 3
Run it yourself
Point your agent at the sample and look for passes that should not pass: an edited test, a new record instead of an updated one.
- 4
Settle upkeep and ownership
Who maintains graders after delivery, who owns them, and whether you get the code or only a hosted copy.
- 5
Check judge agreement
For rubric-scored tasks, how often two judges agree on the same attempt.
Asked before buying an RL environment
Where can I buy RL environments?
From the builders who make them to order. The fourteen companies listed here all work on custom quotes, and some deliver code you run yourself while others host the environment. Scale AI's own page names off-the-shelf sets for tool use and computer use, with task counts. To start for free, open hubs such as Prime Intellect's Environments Hub share environments anyone can reuse. See open-source environments and hubs.
Who are the top RL environment vendors in 2026?
There is no fair ranking yet, because none of them publishes prices or results. Environment specialists include Bespoke Labs, Halluminate and Patronus AI. Large human data providers that also build environments include Scale AI, Turing, Surge AI, Mercor, micro1, Handshake AI, Toloka, Appen and Centific. Prime Intellect and OpenPipe supply open hubs and training tools. The companies page has the source for each.
How much do RL environments sell for?
No company on our list publishes a price. The only public figures come from Epoch AI's interviews, not contracts: in most cases $200 to $2,000 for one task with its grader, with $20,000 rare, and exclusive deals at about 4 to 5 times the shared price. See task costs.
Can a startup buy an RL environment, or is it only for frontier labs?
A startup can. Bespoke Labs, for one, says it sells to newer labs and enterprises as well as frontier labs, and open hubs offer free environments to start from. A small pilot set lets you check grader quality before you commit to a full order.
Which RL environment companies build for coding agents?
Among the companies we list, Appen is the one whose own pages describe a coding set: 500 coding tasks, each with its own automatic verifier. Surge AI, Toloka, Bespoke Labs and Scale AI build environments, rubrics or verifiers to order, and coding is where open environments are most plentiful. No one publishes coding prices. More on builders for coding agents.
How do I check an RL environment's grader before I buy?
Ask for a sample of tasks with their graders that your own engineers can read, the wrong solutions each grader was tested against, and who maintains and owns the graders after delivery. Then run the sample yourself: agents exploit any gap a grader leaves, such as editing a failing test instead of fixing the code. See how graders catch reward hacking.
project brief3 parts
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.