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
Reflection AI expanded its Beam 1.0 501B-A23B lineup with FP8 and NVFP4 precision variants, Tencent introduced the memory-specialized Hy-Memory-30B-A3B, and NVIDIA added a BF16 build of its Nemotron 3.5 Bolt 74B-A9B. Score movements across the top of the leaderboard were mixed but generally small, with Claude Opus 5.5 ticking up to 86.54 while several mid-tier OpenAI and Google models saw modest downward pressure.
Reflection AI Releases Quantized Beam 1.0 Variants
Reflection AI's Beam 1.0 501B-A23B is now available in two additional precision formats: FP8 and NVFP4. These quantized builds of the 501-billion-parameter mixture-of-experts model lower deployment cost and memory footprint while targeting different hardware efficiency profiles, broadening accessibility for large-scale inference.
Tencent
Tencent Debuts Hy-Memory-30B-A3B
Tencent has released Hy-Memory-30B-A3B, a 30-billion-parameter mixture-of-experts model with only 3 billion active parameters per token. The 'Memory' designation signals a focus on long-context or persistent-state capabilities, extending Tencent's Hy model family into specialized memory-augmented territory.
NVIDIA
NVIDIA Adds BF16 Build of Nemotron 3.5 Bolt 74B-A9B
NVIDIA's Nemotron 3.5 Bolt line gains a BF16-precision variant of its 74B-A9B mixture-of-experts model. This joins previously released builds and gives practitioners another numerics option for balancing throughput and accuracy on NVIDIA hardware.
Xiaomi
Claude Opus 5.5 Extends Lead at Top of Leaderboard
Claude Opus 5.5 rose to 86.54 overall, extending its margin over GPT-6.1 Sol (84.06) and GPT-6 Astra (83.87). Xiaomi's MiMo-V2.6-Pro and OpenAI's GPT-5.6 Sol posted the largest drops in the top 15, falling 0.07 and 0.05 points respectively, while Anthropic's Claude Sonnet 5.5 gained 0.03 to reach 82.17.
The catalog
The top of each list. Every list has a full table, sortable and filterable.
Models
By overall score
- 1Claude Opus 5.586.54
- 2GPT-6.1 Sol84.06
- 3GPT-6 Astra83.87
- 4Claude Fable 5.183.17
- 5Claude Sonnet 5.582.17
RL environment companies
By tier, X reputation, benchmarks built
Benchmarks
S tier: the ones frontier labs report
Infrastructure companies
By tier, X reputation, benchmarks built
Browse by domain
All 91 domains and labelsOn the map. Our domains: every company and benchmark sits in one, placed by a quote from its own site or paper.
- Coding
- Enterprise work
- Health and bio
- Computer use
- Math
- Science
- Finance
- Legal
- Media
- Security
- Games
- Robotics
- Multi-domain
- Chip design
Directory labels. Tags other catalogs put on entries. Not domains on the map.
- Agentic
- Multimodal
- Reasoning
- Long-Horizon
- Knowledge
- Safety
- Decision Models
- Multilingual
- Long Context
- Video
The other 67 pages
More pages for domains on the map
- Business Enterprise work
- Biomedicine Health and bio
- Private Codebases Coding
- Software Engineering Coding
- Multi-Domain Multi-domain
- Biology Health and bio
- Robotics Robotics
- Gui Computer use
- Creative Media
- Materials Science
- Creative Media Media
- Chemistry Science
- Spreadsheets Enterprise work
- Security Security
- Browser Computer use
- Text TO Sql Coding
- Games Games
- Image Generation Media
- Video Generation Media
More directory labels
- Retrieval
- Vision
- Multimodal Understanding
- Alignment
- Behavioral
- Tool Use
- Generalization
- Speech
- Spatial Reasoning
- Data
- Factuality
- Instruction Following
- Language
- Document Understanding
- Translation
- Education
- Domain
- Geospatial
- General Knowledge
- Search
- Machine Learning
- ML
- Classification
- Inference
- Question Answering
- Korean
- Aerospace
- Efficiency
- Language-understanding
- Summarization
- Tabular
- Uncategorized
- Writing
- Human-preference
- Mobile
- Structured Output
- Document-ai
- Ocr
- Climate
- Energy
- Meta
- Truthfulness
- Weather
- Self Improvement
- Autonomy
- Openness
- Professional
- Rlhf
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.
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.
Domains
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.
How to read it
A benchmark, in the domain most of its tasks are in. The logo is who built it.
Also covers: a benchmark mostly about another domain that also tests this one.
Lineage: a line from a benchmark to one built on it: its next version, a subset, an extension, or an index that combines it.
Click any star for the line from its own page or paper that put it there, and to light its lineage.
Research
All research posts- Five people, about $300k, synthetic RL environments: Ultramarine 1 Flash beats Claude Fable 5 on 4 of 6 knowledge-work benchmarks
- 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
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.