The RL environment platform landscape in 2026

Source: Dev.to

Why RL environment platforms are emerging

OpenAI, Anthropic, and Meta don’t buy RL environments off the shelf—they build them internally. A TechCrunch investigation reported that Anthropic plans to spend more than $1 billion on RL environments over the next year. OpenAI’s ChatGPT Agent training relies on “UI Gyms,” browser‑based environments that simulate real software at scale. SemiAnalysis notes that major labs maintain distinct procurement strategies, with firms such as Mercor, Surge, and Handshake acting as primary environment and data suppliers.

The market is moving fast. Mercor, one of the largest AI training‑data platforms used by the top five AI labs, acquired Sepal AI in February 2026 to deepen its RL environment capabilities, targeting the intersection of human data, RL environments, and specialized research. TechCrunch highlighted Mercor’s new focus on domain‑specific RL environments for coding, healthcare, and law.

For everyone outside the top labs, building your own environment infrastructure from scratch is almost certainly the wrong move. The engineering cost is high, maintenance is ongoing, and your core competency is likely the agent, not the environment. The platforms below aim to fill that gap.

Platform landscape (2026)

Surge AI – Enterprise RL environments, human‑expert data pipelines

• Partners with OpenAI, Anthropic, Meta, and Google.
• Flagship suite CoreCraft: a large‑scale enterprise simulation with 2,500+ entities and 23 tools, designed to test real‑world agentic capabilities.
• Research shows GPT‑5 and Claude fail over 40 % of agentic tasks in realistic RL environments.
• Trade‑off: Enterprise‑grade pricing; not ideal for smaller teams.

Rise Data Labs – Browser agents, human data pipelines, RL environment curation

• Builds RL training environments focused on human data and AI training pipelines.
• Maintains a curated directory of providers across the ecosystem, offering both a platform and a resource for navigating the broader landscape.
• Well‑suited for teams that aren’t at Surge’s scale but need high‑quality task data.

Mercor – Domain‑specific RL environments, expert data at scale

• Recently acquired Sepal AI to strengthen domain‑specific capabilities (coding, healthcare, law).
• Used by the top five AI labs; leverages a strong human‑expert network for environment and reward design.
• Continues expanding its environment product suite.

Prime Intellect – Research teams, custom environment infrastructure

• Open‑source‑friendly and highly flexible; supports an Environments Hub for bringing your own environments.
• Strong on distributed compute.
• Trade‑off: Onboarding complexity; documentation assumes prior knowledge, making it better for experienced teams.

Mechanize – Coding and software agent tasks

• Purpose‑built for code‑related RL.
• “Replication training” approach: agents recreate implementations from specifications, providing strong reward signals for code tasks.
• Not suitable for browser agents, but valuable for code execution, repo navigation, or terminal interaction.

HUD – General RL, end‑to‑end lifecycle

• A more complete general‑purpose platform covering environment authoring, evaluation, and observability in one place.
• Good for teams that prefer an integrated toolset rather than stitching together separate solutions.
• Performance on browser‑specific tasks lags behind specialized options, but it covers the bases for general RL workflows.

Evaluation considerations

• Match the platform to your task type. A coding‑focused platform won’t meet the needs of browser agents, and vice versa. Specialized platforms excel in their niche but perform poorly outside it.
• Human data integration matters. Platforms that incorporate real human feedback into the reward signal (instead of relying solely on synthetic signals) generally produce agents that generalize better.
• Separate training and evaluation. If you train and evaluate on the same environment, you risk measuring memorization rather than true generalization. Building this separation early is advisable.

If you’ve worked with any of these platforms—or others I haven’t covered—I’d genuinely like to hear what you’ve seen in the comments!