How this directory is researched, sourced, and ranked, and, just as importantly, what we deliberately leave blank. The whole point of rl-list.com is that you can audit every claim.
Procurement teams choosing an RL-environment vendor care about scale, velocity, quality, cost, and data specs. Almost none of that is on the public web, it surfaces only in direct vendor engagement. So we don’t pretend to answer it. Instead, for every company we research the public proxies for those questions, cite each one, and tag how confident we are.
| What a buyer wants to know | The public proxy we source |
|---|---|
| How fast can they scale production | Headcount, headcount growth, open roles, funding |
| Quality & rigor of the work | Researchers on the team, their backgrounds, published papers/benchmarks |
| Will they survive / are they credible | Capital raised, investors, customers, founding year |
| Can we clear security review | SOC 2 / ISO certifications |
| Footprint & jurisdiction | HQ + office locations |
Every non-trivial field on every vendor page carries one of these tags, so you never have to guess how solid a number is:
The numbers frontier labs actually request in an RFI, task/sample counts, unique environments, pass@1 and difficulty, capability and complexity splits, data-type breakdowns, harness and data format, and unit/total pricing, are not on the public web and only come from direct engagement. We do not estimate them, and we do not let an AI “reason toward” a plausible figure. Those fields stay blank on purpose. If you see a number on rl-list.com, it has a source.
We rank only the dedicated, pure-play RL-environment vendors. Three groups are deliberately excluded from the ranking and listed separately for reference, because they aren’t like-for-like comparable: data-labeling incumbents moving into environments (Scale AI, Surge AI, Mercor), execution-infrastructure providers (sandbox/compute layers), and open-source projects. Mixing a $1B labeling incumbent into a list of focused environment startups would mislead more than it informs.
Within the ranked set, order is driven by a transparent formula we call the RL List score, the same calculation applied to every vendor, so the baseline order stays auditable. A small number of vendors we have reviewed in depth are placed editorially; everyone else falls where the score puts them. The score combines:
The score is not a product-quality or endorsement rating, it reflects scale, signal, and how verifiable a vendor’s public record is. A company lower down is often simply earlier-stage or harder to verify, not a weaker product.
Category rankings. The use-case pages (coding, computer use, enterprise workflows) start from the same RL List score, restricted to the vendors that work in that category and renumbered #1..#N, then apply the same allowance for a few editorial placements where we've looked hard at a vendor's fit for that specific use case. So a vendor's category position can differ from its overall rank: a specialist can lead a category it's built for even if a broader vendor sits higher overall. Editorial placements are the exception, not the rule, and incumbents, infrastructure and open-source projects remain unranked everywhere.
Every vendor page shows a “last updated” date, and the directory is re-verified on a rolling basis. This snapshot was last updated 2026-10-05.
Each refresh runs two passes per vendor. A researcher collects everything material from the previous six months, rounds, acquisitions, launches, benchmarks, named customers and key hires, and re-checks the existing record against it. An independent verifier then opens every source to confirm it says what we claim, drops anything it cannot confirm, and looks for what the researcher missed. The vendor page’s “What’s new” timeline and the homepage feed show only items that survived that second pass, each with a date and a link.
Founder names, roles and backgrounds come from the company’s own pages, Y Combinator profiles and credible press. Photos are taken only from a public, professional source where that person is identifiable, such as the company’s team page, a YC profile or a conference speaker page, and we store a small copy rather than linking to someone else’s server. Where we could not find a clear, verifiable headshot we show initials instead, never a guessed or borrowed image.
Benchmarks, datasets, environments and papers a vendor has published are listed with a link to the artifact itself, and a headline result only where the vendor or paper states one. We list them because a team that publishes reproducible benchmarks gives researchers something to inspect before any conversation, which is a different and stronger signal than a capability claim on a marketing page.