RL environments and AI benchmarks, cataloged daily.

Who builds RL environments, who funds them and how models score, rebuilt every morning from public sources. Every placement links to the line on the company's own site that puts it there.

Today in RL

October 19, 2026 was a benchmark-expansion day: 54 new benchmarks landed spanning agentic safety, multi-agent commerce, theory of mind, distributed systems debugging, educational AI, and visual reasoning, while 31 existing leaderboards received updates. No new models, environments, or tools were added. Benchmark coverage expanded broadly across every major frontier lab, with OpenAI, Anthropic, Google, Qwen, and DeepSeek each gaining dozens of additional evaluation appearances.

  • ActBench Targets Agentic Behavioral Safety

    ActBench introduces a self-evolving evaluation of agent behavior risk — specifically whether agents completing benign tasks inadvertently disclose protected data, manipulate unauthorized state, or invoke unauthorized APIs. Each test case pairs a benign task with an adversarial variant, making it one of the first benchmarks to assess behavioral safety from execution trajectories rather than final outputs alone.

  • ACWorld Benchmarks Buyer-Merchant Agent Negotiation

    Agentic Commerce World (ACWorld) introduces a controlled environment where independently controlled Buyer and Merchant agents interact in a shared market, each preserving private objectives. Inspired by 'vibe commerce' — delegating buying and selling goals to AI agents — this benchmark is one of the first to evaluate multi-agent commercial interactions with asymmetric authority and private information.

  • Avalon-ToM-Bench Decomposes Theory of Mind via Asymmetric Information Games

    Avalon-ToM-Bench uses the mechanics of The Resistance: Avalon to evaluate theory of mind in a 2×2 taxonomy — epistemic versus motivational reasoning, first-person versus third-person perspective. Unlike static ToM tests, the benchmark provides fine-grained diagnostic insight into where and how agent mental-state reasoning breaks down.

  • DDBench Opens Distributed-System Debugging as an Evaluation Domain

    DDBench targets a gap left by SWE-style benchmarks: bugs that span processes, nodes, and protocol interactions in distributed systems, where root causes cannot be recovered from source alone. As frontier models cluster in the high-70s on SWE-bench Verified, DDBench pushes evaluation into non-deterministic, multi-process failure scenarios that require reasoning across system boundaries.

  • ClawTrack Scores Both Outcomes and Reasoning Processes

    ClawTrack introduces a dual-assessment framework that simultaneously measures Task Score (what an agent achieves) and Process Score (how it achieves it) in long-horizon multi-step workflows. This addresses a recognized gap in agentic evaluation: benchmarks that judge only final outcomes cannot distinguish reliable reasoning from lucky success.

The catalog

The top of each list. Every list has a full table, sortable and filterable.

Models

By overall score

  1. 1Claude Opus 5.586.56
  2. 2GPT-6.1 Sol84.07
  3. 3GPT-6 Astra83.88
  4. 4Claude Fable 5.183.18
  5. 5Claude Sonnet 5.582.18

All 49 models

RL environment companies

By tier, X reputation, benchmarks built

  1. SSurge AI
  2. SSnorkel AI
  3. SPrime Intellect
  4. SAfterQuery
  5. SMercor

All 70 companies

Benchmarks

S tier: the ones frontier labs report

  1. SHumanity's Last Exam
  2. STerminal-Bench 2.1
  3. SSWE-bench Verified
  4. SAIME 2025
  5. SBrowseComp

All 3,144 benchmarks

Infrastructure companies

By tier, X reputation, benchmarks built

  1. SPrime Intellect
  2. SModal
  3. AE2B
  4. ADaytona
  5. BVeris AI

All 11 companies

Browse by domain

All 91 domains and labels

On the map. Our domains: every company and benchmark sits in one, placed by a quote from its own site or paper.

Directory labels. Tags other catalogs put on entries. Not domains on the map.

The other 67 pages

More pages for domains on the map

More directory labels

The map

Every company we track, placed in the domain it leads with. Click one for the line on its own website that puts it there.

All companies

Every company we track, whatever it sells, placed in the domain it leads with. Companies that cover many domains with none leading sit in the middle.

77companies
15domains
1with no leader

Domains

Multi-domain30cover many domains Coding1313 lead · 32 also cover Infrastructure6runtime, no domain Enterprise work55 lead · 29 also cover Security44 lead · 10 also cover Computer use33 lead · 13 also cover Finance33 lead · 21 also cover Health and bio33 lead · 14 also cover Media33 lead · 9 also cover Robotics22 lead · 9 also cover Science22 lead · 17 also cover Chip design11 lead · 1 also cover Games11 lead · 5 also cover Math11 lead · 9 also cover Legal00 lead · 12 also cover Not yet confirmed5listed, not yet confirmed
Every vendor, with quotes →
Glow · what it leads with selling

The galaxy could not load. See every vendor on the board.

How to read it

A company, in the domain its own site leads with. The logo is its own; the glow is what it mainly sells.

Also covers: a ring at a domain's edge for a company that leads with another domain but also covers this one. A faint line runs to where it leads.

Dots around it: the other things it sells. Hollow while it is still expanding into them.

Click any star for the lines from its own website that put it there.

The map on its own page, with every company listed →

Research

All research posts
  1. Five people, about $300k, synthetic RL environments: Ultramarine 1 Flash beats Claude Fable 5 on 4 of 6 knowledge-work benchmarks
  2. Why expert-data vendors are becoming research labs: 14 studies from four vendors in 40 days
  3. Mercor hired the first authors of LoRA, τ²-bench, SciCode, and APEX-Agents. micro1 is offering up to $4M for a robotics research director
  4. 2,209 agents rewrote a coding harness in Rust in two weeks. The reward was a parity check in a fresh sandbox

In short

RL environments, in short

What are RL environments?
A simulated workplace where an AI agent practises a task (a code repository, a browser, business software) and an automatic grader scores each attempt. Labs train on thousands of these attempts. Bought to improve models. More in the glossary.
Who sells RL environments?
70 companies we track sell RL environments, evals or verifiers, among them AfterQuery, HUD, Mercor, Symbal, Turing, Align AI, BenchFlow and Cua. Each is placed by a line from its own website; see them all on the RL environments page.
What is the difference between an RL environment and a benchmark?
An RL environment is for training: a model practises in it and a grader scores each attempt. A benchmark or eval is for measuring: a fixed, held-out set of tasks with fixed scoring, used to measure and compare models. Never trained on, so the score means something. Used to measure models.
What are the main AI benchmarks for agents and LLMs?
Among the most referenced: Humanity's Last Exam, Terminal-Bench 2.1, SWE-bench Verified, AIME 2025, BrowseComp, Terminal-Bench Hard, GPQA Diamond and GPQA-D. See the full AI benchmarks list.
Where does this data come from?
Every placement is read from the company's own website or the benchmark's own page or paper, quoted and linked, and the catalog is rebuilt daily from public sources. How the data is built.

From the builders

What vendors are posting

The newest post on X from each RL environment vendor we track, refreshed with every build. Scroll the feed for more. How we find their accounts.

  1. Vals AI @ValsAI

  2. Daytona @daytonaio

  3. Snorkel AI @snorkelai

  4. ReasonCore @reasoncoreai

  5. HUD @hud_evals

  6. E2B @e2b

  7. Handshake @joinhandshake

  8. micro1 @micro1_ai

  9. Modal @modal

  10. Taste Labs @tastelabs