If your startup has built reinforcement learning environments, or you run real software that an AI agent could be trained inside, there are buyers for it. AI labs are training agents to complete long tasks in software, and they need realistic places for those agents to practice, with a scorer that decides whether the agent succeeded. Building that is hard, most labs buy it, and the money is real. Read on to find out who's buying, what labs pay per task and per environment, the three ways to sell, and what makes one environment worth more than another.
What labs are buying
An RL environment has three parts (a task space like a replica of a CRM, a codebase or a browser; a set of goals, usually written by people who know the domain; and a verifier that decides whether the agent did the job). Labs pay for all three, and the verifier carries most of the value. A convincing simulation with a scorer that can be gamed teaches an agent to cheat, and buyers have learned to ask how a verifier was tested before they ask anything else. Our guide to RL environment companies maps the vendors already selling into this market.
Who buys and how much they spend
The frontier labs are the buyers, and the budgets are large. In September 2025, The Information reported that leaders at one major lab had discussed spending more than a billion dollars on RL environments over the following year. Vendors at the top of the market describe contracts that run to seven figures a quarter, and smaller labs and newer research groups describe contracts in the 300,000 to 500,000 dollar range per quarter. Established data companies (the ones that already sold human feedback and evaluation) took a large share of that spend first, and environment-native startups founded for this market are taking the rest.
What RL environments sell for
There's no price list, and most work is scoped per deal. Founders and researchers interviewed publicly in late 2025 and early 2026 put the numbers in these ranges:
| What's being sold | Reported price | What it tells you |
|---|---|---|
| A single task with a goal and a verifier | 200 to 2,000 dollars per task in most cases | The unit most contracts are priced in |
| An unusually complex software engineering task | Up to 20,000 dollars, rarely | Depth of verification is what pushes a task to the top of the range |
| A replica of a website or app interface | Around 20,000 dollars per environment | The baseline for a working simulation |
| A high-fidelity replica of a complex product | Around 300,000 dollars | What a full recreation of something like a chat or CRM product costs to buy |
| The same environment sold exclusively | Roughly 4 to 5 times the non-exclusive price | Exclusivity is the biggest lever on price |
Reported prices for RL environments and tasks, from public interviews and reporting, 2025 to 2026.
Two things stand out. First, buyers pay by the task as much as by the environment, so an environment that supports hundreds of well-verified tasks earns far more than a beautiful simulation with ten. Second, exclusivity changes the number more than anything else, and it's the same tradeoff every data owner faces. Our guide to exclusive vs non-exclusive data licenses explains why selling once for more can still pay less than selling several times.
The three ways to sell
Environment builders reach labs in three ways, and most use more than one.
- Sell directly to a lab. The largest contracts run this way. It takes a relationship with a research team, a sample environment they can test, and the ability to build to their spec on their timeline. Expect the lab to test your verifier against reward hacking before anything is signed.
- Build for a data vendor. The established human-data companies sell environments to labs alongside evaluation and feedback work, and many of the environments they deliver are built by smaller teams under contract. You give up margin and the lab relationship, and you get volume and a buyer that already has the contracts.
- Publish on an open hub. Community hubs host thousands of environments with shared verifier libraries and an open API, and they're the fastest way to be seen. They pay mostly in reputation and contributor programs, and labs do hire and license from the people whose environments hold up there.
If you run real software (a product, an internal tool, a workflow your team uses every day), there's a fourth route, and it's often the more valuable one. The record of how that software gets used is training data in its own right. Our guide to selling source code and workflows to AI labs covers how labs price it.
What makes an environment worth more
Buyers price environments on five things:
- How realistic the task space is. A replica that behaves like the real product, edge cases included, is worth more than a clean demo.
- How well the verifier holds up. Buyers test for reward hacking, and a verifier that can't be gamed is the product.
- How many tasks it supports, and how varied they are. Hundreds of distinct, well-specified tasks beat a handful of showcase ones.
- Whether it comes with real data. An environment grounded in real usage records, real tickets or real code teaches an agent things a synthetic one can't.
- Whether you'll maintain it. Software changes, and an environment that tracks the real product stays valuable.
The fourth point is where environment builders and data owners meet. Agents learn what a task looks like from real records before they practice it in a simulation, and labs buy both. Our glossary entry on workflow trajectories explains the unit of data that grounds an environment, and our guide to enterprise operational data covers why labs pay for the real thing.
What to sort out before you sell
Environments carry the same rights questions as any data asset, and three of them come up in every deal:
- Ownership of what you replicated. A replica of someone else's product can raise trademark and terms-of-service questions, and buyers will ask how you built it.
- Real data inside the environment. If the environment uses real records, names and personal details have to be replaced with pseudonyms, and anything under a customer contract stays out.
- Contractor and open-source contributions. Every contributor should have assigned their work to the company, and any open-source code needs a license that allows what you're doing with it.
Our guide to rights-cleared training data explains the standard buyers hold every asset to.
Where to start
If you've built environments, put one in front of a buyer with the verifier tests written up, and price by the task. If what you have is real software and the record of how people use it, start with our free data value assessment (ten questions about your systems, history and industry, about five minutes). It gives you an estimate of what the operating data behind your product could license for, and a scoping call turns that into a real number.
Where Troveo fits
Troveo doesn't build RL environments. We license the real-world data that agent and world-model training runs on (gameplay, task demonstrations, tool use, first-person footage, and the operating records of real companies), sourced from more than 7,000 rights holders and cleared for AI training with documentation for every asset. That's the ground truth an environment is built from and the data that grounds the model before the environment sharpens it. If you've built environments, or you run software whose usage record could train an agent, talk to us about what you have and who's buying it.