The companies that build RL environments, compared

Prime Intellect has the largest public catalog we found, with 2,500+ community environments on its hub, plus the open-source verifiers library, the prime-rl trainer and hosted RL. Its $130M Series A is confirmed, and TechCrunch names Ramp and Zapier as customers. Hub environments are community-built and vary in quality, and it disclosed a sandbox escape affecting eval stacks in August 2026.

Also consider HUD, General Reasoning

HUD is the only vendor with a public usage price: $0.10 per environment-hour on its cloud, with a free SDK tier. The SDK is MIT-licensed, it hosts OSWorld-Verified, and its marketplace lets you buy environments from other vendors. Its SOC 2 claim has no published report, and marketplace environments need your own QA.

Also consider Prime Intellect

Plato documents resettable simulated workplaces (a CRM, Git server, wiki and Ubuntu desktop) with scoring modes, and released Computer Work Val, a 300-world validation set, in October 2026 after a confirmed $16M Series A. The team is about 10 people with no named customers. Runner-up Scale AI publishes 1,000 computer-use tasks across Linux, Windows and macOS, but Meta owns about 49% of it.

Also consider Scale AI

Datacurve’s DeepSWE benchmark is open source (Apache-2.0) and is part of the Artificial Analysis Coding Agent Index, which is rare third-party adoption. Proximal’s FrontierSWE runs 20-hour tasks, and Bespoke Labs contributes to Terminal-Bench 2.0. Datacurve’s headcount fell from about 36 to about 29 between June and October 2026, and Proximal’s $200M revenue claim is unverified.

Also consider Proximal, Bespoke Labs

Turing’s product page names interactive replicas of Jira, Salesforce and Zendesk plus MCP backends with policies written by subject-matter experts, and Epoch AI names it among data providers selling environments. Mercor, through Deeptune, has built hundreds of Slack- and Salesforce-style gyms. Turing’s lab customer claims are self-reported, and Mercor carries a 2026 data breach and pending lawsuits.

Also consider Mercor, Deeptune

Halluminate builds finance diligence environments from anonymized data on real private transactions and has released open Westworld environments. Its $30M Series A (October 2026) is confirmed, and it reports being profitable. It says it works with four of the five top US closed labs without naming any, and one benchmark score differs between its blog and its press coverage.

Collinear’s CWE-bench is included in the Artificial Analysis Cyber Index, and its Simulation Lab offers 100+ runnable environments and 4,000+ tasks. Its trust centre lists SOC 2 and ISO/IEC 42001. Its funding is reported only by an aggregator, and the CWE-bench eval set is private.

OpenEnv is a BSD-3 framework started by Meta’s PyTorch team and Hugging Face and now run by a multi-company technical committee, with 40+ environments and integrations for TRL, SkyRL and Unsloth. General Reasoning’s OpenReward is the vendor alternative, with 430+ hosted environments and the open Open Reward Standard protocol. OpenEnv’s APIs are still early, and General Reasoning also trains its own models.

Also consider General Reasoning

Snorkel AI lists SOC 2 Type II and HIPAA, co-maintains Terminal-Bench and closed a confirmed $350M Series E in September 2026. Scale AI has FedRAMP High, and Huzzle Labs is the only small shop we found with a checkable SOC 2 trust centre. Snorkel’s lab ties come only from a logo wall, and Scale is about 49% owned by Meta.

Also consider Scale AI, Huzzle Labs