Who builds RL environments?
Epoch AI's 2026 interviews describe four kinds of RL environment builders:
| Type | Strength | Trade-off |
|---|---|---|
| Specialist RL environment startups | Focused expertise in a domain such as coding, computer use, math or finance | Smaller teams; capacity and longevity vary |
| Traditional data and labelling providers | Large expert workforces and QA processes, so they can staff projects quickly | Environment engineering depth varies |
| In-house lab teams | Confidentiality, no vendor margin, direct domain expertise | Hiring and management overhead |
| Product companies partnering with labs | Know their own software's edge cases best | Usually limited to their own product |
Epoch AI also reports "substantially more in-housing" recently as labs build their own data teams.
How to evaluate an RL environment company
- Reward-hacking evidence. Ask how they test graders against cheating, and for examples of exploits they found and fixed.
- Pass-rate data. Request pass rates for current models on a sample, to confirm tasks are neither trivial nor impossible.
- Feasibility checks. Are tasks blind-solved by a second expert?
- Fidelity. For app clones, ask for a demo of edge cases and error states; buggy clones teach little.
- Reproducibility. Can you reset and rerun any task and get the same grading?
- Format and framework fit. Will it run in your trainer and sandbox infrastructure?
- Domain expertise. Who writes tasks? Epoch AI quotes a founder: "Domain knowledge and expert level prompting is more important than ML skills for creating tasks."
- Scaling and QA. How is quality kept as volume grows? Builders cited this as the top operational bottleneck.
Pricing and terms to clarify
Reported market prices are a rough guide only: about $200–$2,000 per task, website replicas around $20k and complex product clones around $300k, with exclusive deals about 4–5× non-exclusive (all per Epoch AI). Clarify exclusivity, IP ownership, acceptance criteria, maintenance, and what happens when a grader exploit is found after delivery. More in the RL environment pricing guide.
Quotes based only on task count; no pass-rate data; graders the vendor will not let you inspect; "LLM-judge everything"; no answer when asked how they detect reward hacking.
Get a second opinion on your project
Describe the workflow, target model and success criteria in a short project brief. We review it and, where appropriate, suggest whether to build in-house, adapt an open-source environment, or talk to an independent specialist.
RL environment companies: FAQs
What do RL environment companies sell?
Usually some mix of environments (sandboxes and app clones), tasks with graders, expert-created trajectories, and ongoing quality assurance. Contracts are often priced per task or per environment, with exclusivity raising the price.
How many RL environment startups are there?
We do not know of a reliable public count, and the market changes monthly. Epoch AI and several independent lists track startups in the space, but none is complete. We do not publish a ranked list.
Should we build in-house or buy?
Build in-house if the workflow is core IP, you have engineers with domain knowledge, and you need confidentiality. Buy when you need to scale task creation quickly or need domain experts you cannot hire. Many teams buy a pilot, then decide.
Does RLenvironment.ai build environments?
We are an independent resource, not an agency. We review project briefs and, where it fits, suggest an approach or introduce an independent specialist. We do not guarantee a match.
Sources and further reading
Figures quoted from third-party sources are attributed and may change. RLenvironment.ai is independent and not affiliated with the projects or publishers cited.