Companies building RL environments in 2026: the vendor landscape
The 2026 market for RL environments spans environment-native startups, open ecosystems, human-data companies, and in-house lab teams. Here is how the landscape breaks down.
Companies building reinforcement learning environments in 2026 fall into four groups: environment-native startups like Mechanize, Fleet AI, HUD, Veris AI, Plato, and Bespoke Labs; open ecosystems such as Prime Intellect; human-data companies including Scale AI, Surge AI, Mercor, and Turing that have added environments to their offering; and the frontier labs that build many environments in-house. Beneath all of them sits a supply question that decides whether any environment is worth paying for: who performs the task, who defines what correct looks like, and who grades the agent.
This guide maps the landscape fairly, explains how the four categories differ, and shows where the market is heading. If you are a buyer researching vendors, it is meant to help you choose. For the foundations of how these systems work, start with our hub on reinforcement learning environments.
Why there is a market at all
For years, AI labs bought static labeled datasets. That worked for teaching a model to classify or answer. It does not work for training an agent to complete a job, because a job is a chain of decisions where each step depends on the last. An environment solves this: the agent acts, the environment returns a reward, and the agent improves through repetition.
Two facts explain the spending. First, static data goes stale and the obvious public data is largely used up, so labs now buy environments that generate fresh, proprietary trajectories on demand. Second, the economics are serious. Standard environments run roughly 200 to 2,000 US dollars per task, complex company-replica environments cost around 300,000 US dollars to build, and Anthropic has reportedly discussed spending more than 1 billion US dollars in a single year on RL environments. The broader AI data-labeling and related market crossed 2.3 billion US dollars in 2026, growing about 23 percent year over year, and environments are the fastest-moving slice of it. For how this new data gets generated and why it beats scraped static data, see our companion piece on data generation via RL environments, and our guide to data for AI agents.
The four categories of RL environment companies
1. Environment-native startups
These companies build environments as their core product. They construct the sandbox, the tools, the task interface, and the scoring harness, often for a specific domain or workflow.
- Mechanize focused on automating real work through environments and was acqui-hired by Google DeepMind in September 2026, an early signal of how quickly frontier labs want this capability in-house.
- Fleet AI, HUD, Veris AI, and Plato each build environments that simulate digital work so agents can be trained and measured on realistic tasks.
- Bespoke Labs works on data and environment tooling aimed at making agent training reproducible.
These are strong teams. Their common dependency is the same: an environment only teaches something if a real expert has defined what success means and can grade the result.
2. Open ecosystems
Instead of selling bespoke environments, open ecosystems host shared ones.
- Prime Intellect is often described as the “Hugging Face for RL environments.” Its Environments Hub hosts more than 2,500 environments, and the company raised a 130 million US dollar Series A with reported ARR above 100 million US dollars. The open model lets researchers share and reuse environments rather than rebuild them.
Open hubs lower the cost of starting, but they do not solve the domain-expertise gap. A shared environment for a regulated workflow is only trustworthy if the task and rubric inside it came from a real practitioner.
3. Human-data companies adding environments
These firms built large businesses in data labeling and model evaluation, and are now extending into environments.
- Scale AI and Surge AI bring deep labeling and evaluation operations to the problem.
- Mercor reports roughly 2 billion US dollars in ARR and has been raising near a 20 billion US dollar valuation, with a network of more than 30,000 vetted doctors, lawyers, and PhDs and about 91 percent of revenue from foundation-model labs.
- Turing applies its engineering talent network to agent training and evaluation work.
These companies have scale and expert networks, which is exactly what environments need. The distinction worth probing as a buyer is how deep and how verified the expertise is for your specific domain, because a wrong answer in medicine, law, or finance is costly and a generalist annotator will not catch it.
4. Frontier labs building in-house
The largest labs build many environments themselves, because the environments encode what they most want to improve. The Mechanize acqui-hire shows the pattern: when environment capability becomes strategic, labs buy or build it rather than rent it. Even in-house teams, though, still need outside experts to perform and grade domain tasks they do not employ practitioners for.
Comparison: the four vendor types
| Vendor type | Examples | What they provide | Best for | The gap they still have |
|---|---|---|---|---|
| Environment-native startups | Mechanize, Fleet AI, HUD, Veris AI, Plato, Bespoke Labs | Purpose-built environments and scoring harnesses | Teams that want a custom environment for a specific workflow | Domain experts to perform and grade the task |
| Open ecosystems | Prime Intellect | Shared, reusable environments at scale | Researchers who want to start fast and reuse work | Verified expertise behind each shared environment |
| Human-data companies | Scale AI, Surge AI, Mercor, Turing | Labeling, evaluation, and large talent networks plus environments | Buyers wanting data operations and environments together | Depth and verification of experts for high-stakes domains |
| Frontier labs (in-house) | Large model labs | Proprietary environments tuned to their roadmap | The labs themselves | Outside practitioners for domains they do not staff |
| Expert-supply layer | CleverX | Verified professionals who perform, define, benchmark, and grade | Anyone building or operating an environment who needs real domain judgment | Not an environment builder; it feeds the builders |
Where the market is heading
There are roughly 20 seed-to-Series-A environment startups in 2026, and most observers expect consolidation to three to five leaders by 2030. The forces are familiar: frontier labs want the capability in-house, capital concentrates around the teams with the best data, and buyers eventually prefer fewer, deeper vendors. The September 2026 Mechanize acqui-hire is likely the first of several.
For buyers, the practical takeaway is that the environment-building layer will keep shifting, but the dependency underneath it will not. Whoever builds your environment, its value is capped by the quality of the experts who performed the task, wrote the rubric, and graded the runs. The NIST AI Risk Management Framework makes the same point in risk terms: an agent optimizes exactly what it is rewarded for, so a flawed success definition produces a confidently wrong agent. Academic work on reward design, such as research hosted at arXiv, reinforces how sensitive outcomes are to who sets the reward.
How to choose a vendor
Ask four questions of any environment vendor:
- Who performs the task? Demonstrations should come from someone who actually does the job, not a generalist. See collecting expert demonstrations for AI at quality.
- Who writes the rubric? The success criteria decide what the agent learns. A vague or non-expert rubric trains an agent to look right rather than be right.
- Who grades the runs? Evaluation is where domain judgment matters most. Our guide to evaluating AI agents with domain experts covers how to do this credibly.
- How are the experts verified? In regulated fields, verification by ID, licence, and real experience is what separates usable data from plausible-looking noise.
If the answer to any of these is thin, the environment will run, but it will reward the wrong behavior. That is the single most expensive mistake in agent training.
Where CleverX fits
CleverX is not building environments and does not compete with the companies above. It is the verified-expert-supply layer those builders and labs need. CleverX provides practising professionals across finance and accounting, healthcare, law, engineering and manufacturing, insurance, and operations who perform tasks to create demonstrations, define the rubrics that set what correct looks like, build benchmarks drawn from their real work, and evaluate agent runs inside whichever environment you use. Every expert is verified through government ID, professional licence confirmation, LinkedIn experience matching, and a recorded interview, drawn from a pool of more than 10 million verified professionals. For the broader picture, see domain experts for AI training and evaluation and expert data for AI training.
The environment is the method. The expert judgment inside it is the product. Pick your environment vendor for the engineering, and solve the expert-supply question separately and deliberately, because that is the part that sets the ceiling on everything you train.
Frequently asked questions
Which companies build RL environments in 2026?
The market splits into four groups. Environment-native startups such as Mechanize, Fleet AI, HUD, Veris AI, Plato, and Bespoke Labs build environments as their core product. Open ecosystems like Prime Intellect run shared environment hubs. Human-data companies including Scale AI, Surge AI, Mercor, and Turing have added environments to their labeling and evaluation businesses. And frontier labs build many environments in-house. Underneath all of them sits an expert-supply layer, such as CleverX, that provides the verified professionals who perform and grade the work.
How much does an RL environment cost in 2026?
Pricing ranges from roughly 200 to 2,000 US dollars per task for standard environments, while complex company-replica environments can cost around 300,000 US dollars to build. Anthropic has reportedly discussed spending more than 1 billion US dollars in a single year on RL environments, which signals how central this spend has become to frontier model development.
What is the difference between an environment builder and an expert-supply layer?
An environment builder constructs the simulated world: the sandbox, the tools, the task interface, and the scoring harness. An expert-supply layer provides the verified domain professionals who perform the task to create demonstrations, define the success rubric, and grade agent attempts. The builder supplies the method, and the experts supply the judgment that makes the environment valuable. CleverX operates as the expert-supply layer, not as a competing environment builder.
Is the RL environment market going to consolidate?
Analysts expect it to. There are roughly 20 seed-to-Series-A environment startups in 2026, and most observers expect the field to consolidate to three to five leaders by 2030. The acqui-hire of Mechanize by Google DeepMind in September 2026 is an early example of that consolidation pressure as frontier labs bring environment capability in-house.
Why do AI labs buy environments instead of static labeled data?
Static labeled data teaches a model to recognize patterns, but it cannot teach an agent to complete a multi-step job where each action changes the situation. An environment can, because the agent acts, receives a reward, and retries until it improves. Static data also goes stale, while an environment generates fresh, proprietary trajectories on demand, which is why labs now treat environments as a recurring spend rather than a one-time dataset purchase.
How does CleverX fit into the RL environment landscape?
CleverX is the verified-expert-supply layer that feeds environments rather than a competitor to the companies building them. It provides practising professionals across finance and accounting, healthcare, law, engineering and manufacturing, insurance, and operations who perform tasks, write rubrics, build benchmarks from real work, and evaluate agent runs inside the environments that builders and labs operate.
The companies building RL environments in 2026 are strong, and the field is moving fast. But whichever vendor builds your environment, its value depends on the verified experts who perform, define, and grade the work inside it. That is the layer CleverX is built to supply.