RL environments and AI benchmarks, cataloged daily.
The RL training market as universes you can fly through: every vendor, the RL environment builders, and the benchmarks they built. Each sits in the domain it leads with, and every placement links to the line on its own site or paper that put it there.
Snapshot 2026-10-11, rebuilt daily from public sources.
How the data is built
Every company we track, whatever it sells. Each domain is a realm and each company a star in the one it leads with; generalists orbit the core.
Realms based here
The galaxy could not load. See every vendor on the board.
How to read it
Star, based here: a vendor, in the realm of the domain it leads with. The logo is the vendor's; the glow is what it leads with selling.
Outpost: a ring on a realm's edge for a vendor based elsewhere that also covers that domain. A lane runs home.
Moons: 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.
How to read it
Star, based here: a benchmark, in the realm of the domain most of its tasks are in. The logo is the vendor that built it.
Outpost: a benchmark based in another realm that also tests this domain.
Click any star for the line from its own page or paper that put it there.
Today's signals · 2026-10-11
What changed in the catalog
October 11, 2026 was a quiet leaderboard-churn day. No new entities were added or removed across any dataset — all counts held flat at 3,505 benchmarks, 25,471 models, 1,708 companies, 111 environments, 10 tools, 66 occupations, and 1 research article. All 10 changes were row-count shifts on existing multimodal benchmarks: nine Design Arena leaderboards and the Artificial Analysis Text-to-Image Arena saw minor gains or losses of one to two rows each.
Design Arena Leaderboards Shuffle Rows
Nine Design Arena leaderboards recorded row-count changes today. SVG Design gained the most (+2, now 128 rows), while 3D Design and Data Visualization each shed two rows (down to 159 and 169 respectively). ASCII Art, Image Generation, Logo Design, and Web Design Overall each gained one row. HTML Game Development and UI Components each lost one row. The net effect is effectively neutral churn — human preference votes are flowing in and some older or duplicate rows are being pruned.
Artificial Analysis Text-to-Image Arena Reaches 168 Rows
The Artificial Analysis Text-to-Image Arena leaderboard ticked up by one row to 168 total entries, continuing its steady accumulation of human preference votes across image generation models. This is the only non-Design Arena benchmark to change today.
All Other Datasets Completely Flat
Models, companies, environments, tools, occupations, and research articles all held at their prior counts for a second straight day. Following the score-wipe event on October 10 — which cleared the leaderboard overall scores for 889 models — no score repopulation has occurred yet, leaving the leaderboard in a transitional state.
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.
The catalog
One page per entry
Every environment, benchmark, domain and model has its own page with builder, domains and related work. Sort or filter any table. The snapshot rebuilds daily.
RL environments
Training and eval environments for agents, with builder and domains.
Browse → 3110AI benchmarks
Public eval suites and leaderboards referenced across the ecosystem.
Browse → 93Domains
Coding, computer use, enterprise workflows, and more, each with its entries.
Browse → 0Model leaderboard
Ranked by overall leaderboard score, with creator and rank.
Browse →Domains by size
Ranked by total entries. 95 domains in all.
Show 89 smaller domains ↓ Hide smaller domains ↑
- 07 Reasoning 0 env · 141 bench 141
- 08 Knowledge 0 env · 137 bench 137
- 09 Safety 0 env · 124 bench 124
- 10 Computer Use 23 env · 95 bench 118
- 11 Math 5 env · 97 bench 102
- 12 Science 7 env · 73 bench 80
- 13 Business 0 env · 74 bench 74
- 14 Decision Models 0 env · 74 bench 74
- 15 Finance 6 env · 68 bench 74
- 16 Legal 2 env · 57 bench 59
- 17 Biomedicine 0 env · 58 bench 58
- 18 Multilingual 0 env · 50 bench 50
- 19 Audio 1 env · 44 bench 45
- 20 Cybersecurity 1 env · 40 bench 41
- 21 Long Context 0 env · 40 bench 40
- 22 Video 0 env · 36 bench 36
- 23 Retrieval 0 env · 35 bench 35
- 24 Vision 0 env · 35 bench 35
- 25 Game 0 env · 34 bench 34
- 26 Generalization 0 env · 34 bench 34
- 27 Speech 0 env · 34 bench 34
- 28 Multimodal Understanding 0 env · 33 bench 33
- 29 Behavioral 0 env · 32 bench 32
- 30 Embodied 0 env · 32 bench 32
- 31 Alignment 3 env · 28 bench 31
- 32 Private Codebases 7 env · 23 bench 30
- 33 Spatial Reasoning 0 env · 30 bench 30
- 34 Tool Use 3 env · 27 bench 30
- 35 Software Engineering 0 env · 29 bench 29
- 36 Language 0 env · 26 bench 26
- 37 Data 0 env · 25 bench 25
- 38 Factuality 0 env · 25 bench 25
- 39 Biology 0 env · 24 bench 24
- 40 Document Understanding 0 env · 23 bench 23
- 41 Instruction Following 0 env · 23 bench 23
- 42 Cross-domain 0 env · 22 bench 22
- 43 Education 0 env · 22 bench 22
- 44 Multi-Domain 11 env · 11 bench 22
- 45 Robotics 0 env · 22 bench 22
- 46 Translation 0 env · 22 bench 22
- 47 Gui 0 env · 17 bench 17
- 48 Domain 0 env · 15 bench 15
- 49 Geospatial 0 env · 14 bench 14
- 50 Search 0 env · 13 bench 13
- 51 General Knowledge 0 env · 12 bench 12
- 52 Machine Learning 0 env · 12 bench 12
- 53 ML 7 env · 4 bench 11
- 54 Creative 3 env · 7 bench 10
- 55 Materials 0 env · 10 bench 10
- 56 Classification 0 env · 9 bench 9
- 57 Creative Media 0 env · 9 bench 9
- 58 Inference 0 env · 9 bench 9
- 59 Question Answering 0 env · 9 bench 9
- 60 Korean 0 env · 8 bench 8
- 61 Aerospace 0 env · 7 bench 7
- 62 Efficiency 0 env · 7 bench 7
- 63 Language-understanding 0 env · 7 bench 7
- 64 Summarization 0 env · 7 bench 7
- 65 Tabular 0 env · 6 bench 6
- 66 Uncategorized 0 env · 6 bench 6
- 67 Writing 0 env · 6 bench 6
- 68 Chemistry 0 env · 5 bench 5
- 69 Document-ai 0 env · 5 bench 5
- 70 Human-preference 0 env · 5 bench 5
- 71 Mobile 0 env · 5 bench 5
- 72 Spreadsheets 0 env · 5 bench 5
- 73 Structured Output 0 env · 5 bench 5
- 74 Ocr 0 env · 4 bench 4
- 75 Security 3 env · 1 bench 4
- 76 Browser 3 env · 0 bench 3
- 77 Chip Design 2 env · 1 bench 3
- 78 Climate 0 env · 3 bench 3
- 79 Energy 0 env · 3 bench 3
- 80 Meta 0 env · 3 bench 3
- 81 Multi-domain 2 env · 1 bench 3
- 82 Text TO Sql 0 env · 3 bench 3
- 83 Truthfulness 0 env · 3 bench 3
- 84 Weather 0 env · 3 bench 3
- 85 Self Improvement 0 env · 2 bench 2
- 86 Tool-use 1 env · 1 bench 2
- 87 Autonomy 0 env · 1 bench 1
- 88 Chart Reasoning 0 env · 1 bench 1
- 89 Games 1 env · 0 bench 1
- 90 Image Generation 0 env · 1 bench 1
- 91 Openness 0 env · 1 bench 1
- 92 Other 0 env · 1 bench 1
- 93 Professional 0 env · 1 bench 1
- 94 Rlhf 1 env · 0 bench 1
- 95 Video Generation 0 env · 1 bench 1
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?
- 53 companies we track sell RL environments, among them AfterQuery, HUD, Mercor, Toloka, Align AI, BenchFlow, Cua and Datacurve. Each is placed by a line from its own website; see them all on the RL environments market map.
- 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, OSWorld 2.0, SWE-bench Pro, Terminal-Bench 3.0, APEX-Agents, APEX and DeepSWE. 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.
- Why expert-data vendors are becoming research labs: 14 studies from four vendors in 40 days
- Mercor hired the first authors of LoRA, τ²-bench, SciCode, and APEX-Agents. micro1 is offering up to $4M for a robotics research director
- 2,209 agents rewrote a coding harness in Rust in two weeks. The reward was a parity check in a fresh sandbox