RL environments market map

Who sells what in RL training: every vendor we track, sorted by what it sells and the domains it covers. Each placement is read from the vendor's own website and quoted.

82 vendors · read November 14, 2026 · click a vendor to see the quotes

The vendor universe

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.

What each company sells

Pick one to see who sells it. A company can sell into several: Scale AI sells RL environments and data.

What each term means →

Showing all 82 vendors

Multi-domain 28

Generalists: vendors covering three or more domains without one they lead with. Their specific domains are still listed.

  • Turing RL envsData

    Turing supplies frontier AI labs with expert human data, datasets, RL environments and evals, and provides enterprises with forward-deployed engineers and an AI agent control plane.

    Why it's in Multi-domain: it covers 8 domains with none leading

    • Coding“Software engineeringFor the software development lifecycle.” source ↗
    • Enterprise work“Enterprise knowledge workFor finance, medical, legal, and more.” source ↗
    • Finance“For finance, medical, legal, and more.” source ↗
    • Legal“For finance, medical, legal, and more.” source ↗
    • Health and bio“For physics, chemistry, math, biology, and more.” source ↗
    • Science“For physics, chemistry, math, biology, and more.” source ↗
    • Math“For physics, chemistry, math, biology, and more.” source ↗
    • Security“CyberStrike: New cybersecurity benchmark for frontier agents” source ↗

    What it sells

    • Human data“Request dataset samples or join our global expert network to advance superintelligence.” source ↗
    • Datasets“SciCode++: Scientific coding data for frontier models” source ↗
    • RL environments expanding“27+ RL ENVIRONMENTS” source ↗
    • Evals expanding“Introducing CEO Bench: Evaluating frontier agents on real company work” source ↗
    • Enterprise AI services“We embed forward-deployed engineers directly in your organization to build and operate AI agents.” source ↗
  • Abundant RL envsDataInfra

    Abundant builds RL environments, datasets and custom evals for frontier AI labs, focused on coding, sciences, math and robotics.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Coding“Turning PRs from any GitHub repo into SWE-bench style RL environments.” source ↗
    • Science“Deep in coding, sciences, engineering, RSI, alignment, and robotics” source ↗
    • Robotics“Deep in coding, sciences, engineering, RSI, alignment, and robotics” source ↗
    • Math“TheoremSmith, a pipeline for generating synthetic formal verification environments.” source ↗

    What it sells

    • RL environments“We specialize in creating environments and datasets for RL by leveraging our experience in simulation and model training.” source ↗
    • Datasets“We've created 6 datasets spanning the most popular programming languages. Each dataset contains 1000 validated tasks harvested from popular GitHub repos.” source ↗
    • Evals“Custom evals & post-training” source ↗
    • Training platforms expanding“Foundry turns your own data and workflows into RL environments. Train and serve the result on Fireworks AI.” source ↗
  • AfterQuery RL envsData

    AfterQuery is an expert data lab that sells SFT, RLHF and rubric-based RL data, tool-calling and computer-use RL environments, off-the-shelf datasets and custom evals built by a network of verified professionals.

    Why it's in Multi-domain: it covers 7 domains with none leading

    • Coding“Spans expert-written code, test cases, and debugging traces that teach models to write production-quality software” source ↗
    • Computer use“Human-demonstrated interactions across browser and desktop environments — teaching models to navigate and operate software end-to-end.” source ↗
    • Enterprise work“NVIDIA publicly used AfterQuery’s Off-The-Shelf Office Agent Training Dataset to improve Nemotron 3 Ultra on GDPval” source ↗
    • Finance“AfterQuery and Hebbia post-trained the NVIDIA Nemotron 3.5 Super VL open model on expert-written public credit tasks.” source ↗
    • Legal“Building a Frontier Legal Evaluation in Partnership with Legora” source ↗
    • Health and bio“verified practitioners across medicine, law, finance, engineering, and more” source ↗
    • Media“Video editors in Premiere, DaVinci Resolve, and After Effects to evaluate video generation and demonstrate timeline workflows for frontier AI labs.” source ↗

    What it sells

    • RL environments“Provides custom RL environments built on top of real APIs, MCP servers, and developer tools” source ↗
    • Datasets“Offers ready-to-deploy datasets across high-demand capability areas” source ↗
    • Human data“Draws on 300,000+ verified practitioners across medicine, law, finance, engineering, and more” source ↗
    • Evals“Tailors evaluation suites and training datasets to your specific capability targets” source ↗
  • Align AI RL envsData

    Align AI sells a runtime verification layer for AI agents, along with domain-specific RL environments, a ground-truth engine, evaluation frameworks and a licensed expert network for regulated, high-stakes domains.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Legal“We translate laws, regulations, and internal policies into machine-checkable constraints” source ↗
    • Enterprise work“We ground verification in enterprise knowledge bases and systems of record” source ↗
    • Health and bio“clinical reasoning, underwriting intuition, decisions that can't be looked up.” source ↗
    • Finance“clinical reasoning, underwriting intuition, decisions that can't be looked up.” source ↗

    What it sells

    • Agent safety and verification“Verification runs inside the runtime, before every step.” source ↗
    • RL environments“Purpose-built reinforcement learning environments for long, complex tasks” source ↗
    • Human data“we source, verify, and manage hundreds of licensed domain experts — clinicians, scientists, operators — within days, not months.” source ↗
    • Evals“We compose these sources into verification layers and evaluation frameworks” source ↗
  • Chakra Labs RL envsData

    Chakra Labs sells expert-authored tasks with graders and pixel-perfect cloned software environments for training and evaluating AI agents in finance, legal, safety and computer use.

    Why it's in Multi-domain: it covers 5 domains with none leading

    • Finance“Finance tasks authored by private equity, hedge fund, and investment banking professionals, spanning modeling, diligence, and interactive dashboards.” source ↗
    • Legal“Legal tasks authored by seasoned practicing counsel in a wide range of practice areas.” source ↗
    • Security“red-teaming agents with direct and indirect prompt injection inside cloned enterprise environments” source ↗
    • Computer use“Computer-use tasks authored across long-horizon, multi-step workflows, requiring agents to cross between GUI interaction and tool-call execution within a single trajectory.” source ↗
    • Enterprise work“red-teaming agents with direct and indirect prompt injection inside cloned enterprise environments” source ↗

    What it sells

    • RL environments“We build expert-authored tasks and pixel-perfect environments to train models and evaluate their real-world capability.” source ↗
    • Evals“The matching task set is delivered, ready to run in your evaluation pipeline.” source ↗
    • Human data“Each domain runs as its own factory: tasks written by experts, graders built for real work.” source ↗
    • Datasets“Bespoke Datasets” source ↗
  • Handshake RL envsData

    Handshake AI sells reinforcement learning environments, benchmarks and expert human data for training and evaluating frontier and enterprise AI models, alongside fine-tuning help for enterprises.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Finance“ATLAS-Finance puts models through 100 tasks spanning complex deal work like M&A advisory, LBO modeling, structured finance, and valuations.” source ↗
    • Health and bio“adapting Nemotron 3.5 Super VL to understand visual artifacts in drug discovery and biotech tasks.” source ↗
    • Coding“Coding Agents Build for the Grader They Imagine, Not the User: Speculative Reward Hacking in DeepSWE” source ↗
    • Enterprise work“a workbook whose formulas must tie across tabs, a deck whose charts must match the underlying data, or an email that must actually have been sent.” source ↗

    What it sells

    • RL environments“Handshake AI is partnering with Fireworks AI to bring enterprises the reinforcement learning environments trusted by frontier labs.” source ↗
    • Evals“ATLAS-Finance is a Handshake AI benchmark recently added to the Fireworks Specialized Intelligence Index.” source ↗
    • Human data“Exploring why human data is central to trustworthy AI and why ethical, careful stewardship is essential for the next era of model training.” source ↗
    • Enterprise AI services expanding“Handshake AI helps enterprises fine-tune OSS models for their domains.” source ↗
  • HUD RL envsDataInfra

    HUD sells a platform and SDK for building, running, QA-ing and training on RL environments and evals, with hosted environment runtime and a DataVendor marketplace where builders sell environments and training data to AI labs.

    Why it's in Multi-domain: it covers 5 domains with none leading

    • Coding“Codinghud-evals” source ↗
    • Computer use“Computer usehud-evals” source ↗
    • Robotics“Roboticshud-evals” source ↗
    • Enterprise work“How UiPath benchmarks enterprise agents against every frontier model” source ↗
    • Games“ARC-AGI-3hud-evals” source ↗

    What it sells

    • RL environments“The platform for building RLenvironments” source ↗
    • Evals“Encode your expertise into environments to train and evaluate models,and create the post-training data that aligns AI to your work.” source ↗
    • Datasets expanding“List them on DataVendor, where AI labs buy training and evaluation data.” source ↗
    • Sandboxes and compute“Modal sandboxes are billed by the second for the CPU cores, memory, and GPUs they request” source ↗
  • Huzzle Labs RL envsData

    Huzzle Labs builds long-horizon RL environments for coding, computer use and enterprise workflows for frontier AI labs, backed by a 300k+ expert network, and also offers enterprise custom evals and models plus an insurance benchmark.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Coding“Multi-file repos with build, run and test loops.” source ↗
    • Computer use“Full desktop and browser control, judged on the end state of long, multi-step tasks.” source ↗
    • Enterprise work“CRMs, spreadsheets, ticketing and finance — the real work companies run, with custom graders on your data.” source ↗
    • Finance“A benchmark for how language models handle real insurance work.” source ↗

    What it sells

    • RL environments“We build **RL environments** for frontier AI labs” source ↗
    • Human data“Our **AI recruiter** sources vetted specialists for any domain, on demand.” source ↗
    • Evals expanding“We are building benchmarks that test real professional work, starting with InsureBench” source ↗
    • Enterprise AI services expanding“the **evaluations and custom models** that put AI to work inside the enterprise.” source ↗
  • micro1 RL envsData

    micro1 is an AI data lab selling expert human data, RL environments (Realm), contextual agent evaluations (Cortex) and robotics data to frontier labs and enterprises.

    Why it's in Multi-domain: it covers 6 domains with none leading

    • Robotics“High-fidelity real-world robotics data to train the next generation of embodied systems” source ↗
    • Health and bio“The micro1 Medical Index brings our medical benchmarks together in one place.” source ↗
    • Coding“Source code documentation, engineering workflows, bug tracking, product planning, and technical knowledge bases” source ↗
    • Finance“Financial processes, SOPs, reconciliations, approvals, reporting workflows, and operational documentation” source ↗
    • Legal“Policies, contract workflows, review processes, governance documentation, and compliance procedures” source ↗
    • Enterprise work“CRM data, sales processes, playbooks, customer lifecycle documentation, and revenue operations” source ↗

    What it sells

    • RL environments“RL environments that mirror real-world scenarios to generate world-class human data for agentic actions and elevate model reasoning” source ↗
    • Human data“We're building the infrastructure for advancing intelligence through expert human data, real-world training environments, and contextual evaluations” source ↗
    • Evals“The contextual evaluation platform for improving AI agent performance in production” source ↗
    • Datasets expanding“micro1 pays companies to license the valuable operational data they already generate, helping advance the next generation of AI” source ↗
  • Prime Intellect RL envsInfra

    Prime Intellect sells an integrated stack for training and deploying custom models: hosted RL training, evaluations, an open environments hub, microVM sandboxes, inference and GPU compute.

    Why it's in Multi-domain: it covers 8 domains with none leading

    • Coding“DeepSWE long-horizon software engineering tasks from Harbor Hub.” source ↗
    • Finance“FINANCE AUTOMATION” source ↗
    • Science“SCIENTIFICDISCOVERIES” source ↗
    • Chip design“CHIP DESIGNAUTOMATION” source ↗
    • Games“CodingData & MLGamesMath & ReasoningMultimodalScience & MedicineSearch & Tool Use” source ↗
    • Math“Math & Reasoning” source ↗
    • Health and bio“Science & Medicine” source ↗
    • Media“Multimodal” source ↗

    What it sells

    • Training platforms“Train models with reinforcement learning on managed Prime-RL, our open-source trainer, using hosted compute and sandboxes.” source ↗
    • Evals“Create hosted evaluations for you to benchmark the performance of your models.” source ↗
    • RL environments“Build RL environments with our open-source library, or start from 2,500+ community environments on the Hub.” source ↗
    • Sandboxes and compute“Scale thousands of microVM sandboxes with native Docker and full RL support.” source ↗
  • ReasonCore RL envsData

    ReasonCore AI sells expert-verified reasoning datasets, sandboxed RL environments and adaptive evals-as-a-service across scientific, coding, financial and spatial domains to frontier AI labs and enterprises.

    Why it's in Multi-domain: it covers 7 domains with none leading

    • Science“Comprehensive STEM training data spanning physics, materials science, chemistry, biology, and mathematics.” source ↗
    • Math“SciCode, FrontierScience, FrontierMath, CritPt Benchmark Training Sets” source ↗
    • Health and bio“Physics · Chem · Bio · Math” source ↗
    • Coding“including data targeting SWE-Bench (real-world GitHub bug resolution and patch generation) and Terminal-Bench” source ↗
    • Finance“Specialized economic and financial valuation datasets covering macroeconomic indicators, GDP modeling, financial analysis, and fraud detection.” source ↗
    • Robotics“Physical AI systems — autonomous vehicles, legged robots, industrial arms, aerial drones, and embodied agents of all kinds” source ↗
    • Enterprise work“These environments are modeled on enterprise software at its most demanding — operational dashboards, financial platforms, infrastructure tooling, compliance workflows” source ↗

    What it sells

    • Datasets“We produce exclusive and off-the-shelf (OTS) datasets built around multi-step hierarchical reasoning, tool-augmented workflows, and domain-specific problems” source ↗
    • RL environments“We design and maintain sandboxed RL/RLVR environments that let models learn through interaction with realistic, high-stakes systems.” source ↗
    • Evals“We develop living, adaptive evals that evolve as model capabilities advance.” source ↗
    • Human data“PhD-level domain experts guide every stage — from problem design to final verification — with authorship models tailored to your pipeline.” source ↗
  • Refresh RL envsData

    Refresh builds simulated software environments with verifiable rewards (computer-use software clones, Blender 3D animation, coding tasks) used by AI labs and enterprises to train and evaluate agents.

    Why it's in Multi-domain: it covers 5 domains with none leading

    • Coding“Software engineering work that targets GDPval capabilities” source ↗
    • Computer use“High-fidelity clones of full software suites, wiring together multiple connected applications, from EHRs to enterprise apps and beyond.” source ↗
    • Enterprise work“from EHRs to enterprise apps and beyond.” source ↗
    • Health and bio“from EHRs to enterprise apps and beyond.” source ↗
    • Media“Agents working inside Blender, from building scenes and materials to rigging characters, keyframing animation, and running cloth and physics simulations.” source ↗

    What it sells

    • RL environments“Refresh builds simulated worlds for coding and computer use, with verifiable rewards that move models forward.” source ↗
    • Evals“Refresh builds these environments as both evals, to objectively benchmark how well frontier models perform real computer work” source ↗
    • Human data“Refresh builds environments by partnering with expert engineers in our network to source real tasks that break the leading models.” source ↗
    • Datasets“Email contact@refresh.dev to request environments or discuss a bespoke dataset” source ↗
  • Rise Data Labs RL envsData

    Rise Data Labs supplies expert human data, RL environments for tool-calling and computer use, and custom evaluations to AI labs and enterprise teams, drawing on a network of US professionals.

    Why it's in Multi-domain: it covers 6 domains with none leading

    • Computer use“Desktop and browser tasks with expert trajectories for navigation, form completion, document handling, and multi-step workflows.” source ↗
    • Health and bio“Specialists in medicine, law, finance, engineering, and other professional fields create tasks, review model outputs, and build rubrics around domain requirements.” source ↗
    • Legal“Medicine Law Finance Engineering” source ↗
    • Finance“Medicine Law Finance Engineering” source ↗
    • Coding“We build environments around APIs, MCP servers, and developer tools.” source ↗
    • Enterprise work“Find a customer record and prepare a verified account update.” source ↗

    What it sells

    • Human data“Source, vet, and match specialists from our network of 500k+ US professionals.” source ↗
    • RL environments“RL environments built around APIs, MCP servers, and developer tools, with expert examples of tool use and error recovery.” source ↗
    • Evals“Find where models fail and why. Experts build rubrics, review results, and use verifiers to check outcomes against the task’s requirements.” source ↗
    • Datasets“SFT RLHF Off-the-shelf data” source ↗
  • Scale AI RL envsData

    Scale AI sells expert-sourced AI training data, RLHF, evaluation and red-teaming for model builders, plus enterprise and government AI applications.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Robotics“Fuelling the next generation of robotic foundation models with real-world training data.” source ↗
    • Health and bio“Reducing physician cognitive load by turning complex patient records into clinical intelligence.” source ↗
    • Security“Use prompt injection techniques to find vulnerabilities.” source ↗
    • Media“Scale Image” source ↗

    What it sells

    • Human data“We source contributors with precision (25% have advanced degrees) and deliver at the bar frontier AI demands.” source ↗
    • Evals“Evaluate your model against a set of complex and diverse prompts to find weak points.” source ↗
    • Agent safety and verification“Use prompt injection techniques to find vulnerabilities.” source ↗
    • Enterprise AI services“We find the right use case, build the system, and own the outcome.” source ↗
  • Surge AI RL envsData

    Surge AI sells expert human data (RLHF, SFT, rubrics, human evaluation), RL environments with verifiers, benchmarks and off-the-shelf datasets to frontier AI labs.

    Why it's in Multi-domain: it covers 7 domains with none leading

    • Coding“spanning RL environments, coding, and the core capabilities every model needs.” source ↗
    • Computer use“teaching them to use computers, navigate the web, and reason for the first time.” source ↗
    • Finance“DAYJOB: TaxA benchmark for AI tax agents” source ↗
    • Legal“doctors, lawyers, investment bankers, Fields Medalists” source ↗
    • Health and bio“doctors, lawyers, investment bankers, Fields Medalists, Harvard professors” source ↗
    • Math“Fields Medalists” source ↗
    • Media“We teach AI to see, watch, and hear — teaching them to generate and understand images, audio, and video.” source ↗

    What it sells

    • Human data“We look for the most brilliant minds on the planet in every domain – doctors, lawyers, investment bankers, Fields Medalists, Harvard professors” source ↗
    • RL environments“We create rich, complex RL environments that challenge agentic models in novel ways, and design verifiers that reward their behavior.” source ↗
    • Evals“DAYJOB: TaxA benchmark for AI tax agents” source ↗
    • Datasets“Our off-the-shelf datasets are the same: thousands of hours of expert reasoning, pre-built and ready to use today” source ↗
  • Theta RL envsDataInfra

    Theta builds custom RL environments from client and expert data to train specialized agents for frontier labs and enterprises, plus a computer-use benchmark (CUB) and forward-deployed agent solutions.

    Why it's in Multi-domain: it covers 6 domains with none leading

    • Finance“Generate DCF Model” source ↗
    • Coding“Review Pull Request” source ↗
    • Health and bio“Analyze Patient Labs” source ↗
    • Legal“Redline Contract” source ↗
    • Enterprise work“Whether it's back office work, sophisticated financial modeling, or refactoring a codebase” source ↗
    • Computer use“CUB is the most diverse and representative benchmark for evaluating computer and browser use agents.” source ↗

    What it sells

    • RL environments“We synthesize your first party data and our human expert data into environments that reflect the real goals, workflows, and outcomes of your business.” source ↗
    • Evals“We're excited to announce CUB, a first-of-its-kind benchmark for computer and browser use agents.” source ↗
    • Enterprise AI services“We work side by side with your team to embed your agent workforce, with rapid refinement to ensure reliability and alignment.” source ↗
    • Sandboxes and compute expanding“we're also introducing our computer environment infrastructure for effective evaluation and post-training at scale.” source ↗
  • Vetto AI RL envsData

    Vetto AI is a post-training data research partner that builds RL environments, verifiers, evaluation benchmarks, expert human data and off-the-shelf datasets for frontier AI labs.

    Why it's in Multi-domain: it covers 6 domains with none leading

    • Coding“Pre-built datasets spanning RL environments, coding, and core capabilities.” source ↗
    • Science“from practitioners at the frontier of the sciences, medicine, law, finance, engineering, and the humanities.” source ↗
    • Health and bio“from practitioners at the frontier of the sciences, medicine, law, finance, engineering, and the humanities.” source ↗
    • Legal“from practitioners at the frontier of the sciences, medicine, law, finance, engineering, and the humanities.” source ↗
    • Finance“from practitioners at the frontier of the sciences, medicine, law, finance, engineering, and the humanities.” source ↗
    • Media“Data that teaches models to see, hear, and watch, understanding and generating across images, audio, and video.” source ↗

    What it sells

    • RL environments“Interactive environments where agents take on open-ended tasks and make real decisions, with verifiers built to reward the behavior that matters.” source ↗
    • Evals“Original tasks graded by a private verifier, comparing frontier models against failure modes, so labs find the next gap to close.” source ↗
    • Human data“Human judgment for what formulas miss: whether an answer is sound, useful, safe, and actually right, not just plausible.” source ↗
    • Datasets“Pre-built datasets spanning RL environments, coding, and core capabilities. Available immediately, no custom build required.” source ↗
  • Anthromind RL envsData

    Anthromind sells expert-generated, fully managed data curation for post-training, RLHF and custom evaluations of AI models.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Legal“It can be domain specific (legal, financial etc.) or use case specific (coding, RAG, tool use etc.).” source ↗
    • Finance“It can be domain specific (legal, financial etc.) or use case specific (coding, RAG, tool use etc.).” source ↗
    • Coding“It can be domain specific (legal, financial etc.) or use case specific (coding, RAG, tool use etc.).” source ↗
    • Health and bio“enable your generative AI models to excel in unique scholarly fields, from humanities to biomedical research.” source ↗

    What it sells

    • Human data“Fully managed, large-scale data curation with dedicated expert teams.” source ↗
    • Evals“Verify your specific applications, test your ideas, and ensure your business goals are met through evaluations designed specifically for your requirements.” source ↗
    • Datasets“Leverage curated datasets tailored to niche or emerging research areas.” source ↗
  • Aptura RL envsData

    Aptura provides expert human data, curated datasets and RL environments to train and evaluate AI on finance, healthcare and legal tasks.

    Why it's in Multi-domain: it covers 3 domains with none leading

    • Finance“Agents complete tasks in finance spreadsheets, medical records, legal documents—with measurable outcomes.” source ↗
    • Health and bio“I review oncology cases from home to help train AI models that support cancer diagnosis.” source ↗
    • Legal“We help train and evaluate AI to master real-world tasks in finance, healthcare, and legal.” source ↗

    What it sells

    • Human data“Physicians, attorneys, and investment analysts with years of high-stakes experience, supported by a platform built for structured, high-quality data collection at scale.” source ↗
    • Datasets“High-quality datasets. Domain-specific data that reflects the real complexity of regulated work.” source ↗
    • RL environments“Rich environments loaded with real data—verifiable rewards, custom tooling, data licensing support.” source ↗
  • BenchFlow RL envsInfra

    BenchFlow is a frontier environment lab that builds open-source RL environments and agent benchmarks (SkillsBench, ClawsBench, Robo Use, FrontierPhysics) plus a runtime for running environments and evals.

    Why it's in Multi-domain: it covers 5 domains with none leading

    • Enterprise work“Five wire-compatible workplaces. Capability and safety, scored separately.” source ↗
    • Robotics“Coding agents drive simulated robots from the shell. 341 tasks, 14 suites.” source ↗
    • Science“Evaluating agents for end-to-end frontier physics research.” source ↗
    • Computer use“We build the environments AI agents need to learn and be evaluated on real computer work, not static prompts.” source ↗
    • Coding“BenchFlow also reads the `task.toml` layout that Terminal-Bench and many other benchmarks use.” source ↗

    What it sells

    • RL environments“We build stateful environments with services, files, tools, verifiers, traces, and replay, so agents can learn from whole workflows.” source ↗
    • Evals“Do procedural skills make agents better at real work? 87 tasks, 8 domains.” source ↗
    • Sandboxes and compute“Run a real task locally: Docker, an ACP agent, an independent verifier.” source ↗
  • Collinear RL envsData

    Collinear sells Simulation Lab RL environments, off-the-shelf tasks, verifier-graded agent training data and benchmarks to AI labs for training and evaluating agents.

    Why it's in Multi-domain: it covers 5 domains with none leading

    • Coding“Software engineering, computer use, and long-horizon reasoning” source ↗
    • Computer use“From GUIs and APIs to MCP, CLI, GitHub, Jira, ServiceNow, and Salesforce” source ↗
    • Security“Cybersecurity, finance, and customer service” source ↗
    • Finance“Expert-calibrated where correctness is hard. Cybersecurity, finance, and customer service” source ↗
    • Enterprise work“Stateful sandboxes / clones of enterprise tools with state persistence across turns.” source ↗

    What it sells

    • RL environments“100+ runnable environments built across advanced agent capabilities” source ↗
    • Evals“Benchmark Diagnose failure, don’t just rank models. Including CWE-bench, FrontierCUA, MCP-Atlas, and OSWorld-Verified” source ↗
    • Datasets“500B+ tokens of agent training data generated” source ↗
  • Mercor RL envsData

    Mercor runs a network of paid professional experts and supplies human data, evaluations/benchmarks and licensed enterprise datasets to frontier AI labs and enterprises.

    Why it's in Multi-domain: it covers 6 domains with none leading

    • Legal“Attorney-built benchmark of 1,749 real legal tasks across 24 practice areas, graded on finished work products like memos and redlines.” source ↗
    • Finance“Long-horizon, cross-application tasks in professional accounting” source ↗
    • Coding“Measures AI performance on real-world software engineering tasks, from bug fixes to feature builds.” source ↗
    • Science“evaluate agents on real mechanical, electrical, civil, and structural engineering work” source ↗
    • Health and bio“Clinical notes, claims, intake forms” source ↗
    • Enterprise work“Your internal data captures how organizations actually operate, which is exactly what labs need to build AI that can handle real enterprise tasks.” source ↗

    What it sells

    • Human data“Join our network of experts solving AI challenges” source ↗
    • Evals“Mercor collaborates with enterprises to build benchmarks that measure AI capabilities relevant to their specific industries, products, and workflows.” source ↗
    • Datasets expanding“The data your business generates every day is exactly what frontier labs need. We license it, package it, and pay you for it.” source ↗
  • Snorkel AI RL envsData

    Snorkel AI sells expert-built specialized training data, RL environments and evaluation benchmarks to frontier AI labs and enterprise AI teams.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Coding“Evaluates terminal agents on continuously updated software-engineering tasks.” source ↗
    • Computer use“Evaluates computer-use agents on 108 long-horizon workflows.” source ↗
    • Enterprise work“Evaluating AI agents on long-horizon, economically valuable professional workflows with verifiable outcomes.” source ↗
    • Finance“Benchmarking Agents in Insurance Underwriting” source ↗

    What it sells

    • Datasets“Snorkel helps frontier labs and AI teams develop specialized training data and environments that set their models and agents apart.” source ↗
    • RL environments“develop specialized training data and environments that set their models and agents apart.” source ↗
    • Evals“A benchmark for evaluating computer-use agents on long-horizon, real-world workflows” source ↗
  • Pareto RL envsData

    Pareto sells human-generated training data and evaluation work for AI labs, sourced from an expert network.

    What it sells

    • Human data“constantly advancing the frontier of human-based training signal through the products we build and the research that drives them.” source ↗
    • Evals“Pareto Samples is how Pareto shares evidence from its data and evaluation work with the research teams it works with.” source ↗
  • Patronus AI RL envsData

    Patronus AI is a behavioral research lab that builds simulated work environments for AI agents to practise in and creates training data to correct agent failures, alongside evaluation research.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Coding“Ship a feature in code no one documented.” source ↗
    • Enterprise work“Knowledge Work→Make one decision from sources that disagree.” source ↗
    • Science“Identify and explore the untested question.” source ↗
    • Computer use“Get work done in software built for people.” source ↗

    What it sells

    • RL environments“We build worlds for agents to practice” source ↗
    • Datasets“When an agent misses the point or cuts corners, we figure out why. Then we build the training data to correct it.” source ↗
  • Preference Model RL envs

    Preference Model builds reward-hacking-resistant RL environments for frontier labs to post-train models, with tasks spanning GPU kernel coding, games, math post-training and data curation.

    Why it's in Multi-domain: it covers 3 domains with none leading

    • Coding“Write CUDA kernels for the sliding-window attention layer of Inkling, Thinking Machines Lab's open-weights model, which no public kernel covers.” source ↗
    • Games“Play the real Slay the Spire through game tools alone: pick a route through each act's map, win turn-based card fights” source ↗
    • Math“Post-train Qwen3-4B, an open reasoning model, to solve competition math problems with far less reasoning, without losing accuracy.” source ↗

    What it sells

    • RL environments“That's why we're building RL environments for training capable, better-aligned superintelligences.” source ↗
  • Vals AI RL envs

    Vals AI is an independent third-party evaluator that benchmarks frontier AI models on economically valuable professional tasks and frontier risks, publishing the Vals Index and leaderboards.

    Why it's in Multi-domain: it covers 6 domains with none leading

    • Finance“We benchmark the world's leading AI models on economically valuable tasks such as finance, software” source ↗
    • Coding“#1 of 66 on ProgramBench” source ↗
    • Security“frontier risk like cybersecurity, recursive self improvement and mental health” source ↗
    • Legal“#4 of 36 on the Legal Index” source ↗
    • Health and bio“#1 of 31 on BioMysteryBench” source ↗
    • Science“#2 of 24 on the Vals Science Index” source ↗

    What it sells

    • Evals“We run all of our own evaluations and create many of our benchmarks in-house.” source ↗
  • Vmax RL envs

    Vmax sells Campaign, self-play task generation infrastructure that turns a customer's proprietary data and evals into new RL environments and refines agents on them.

    Why it's in Multi-domain: it covers 4 domains with none leading

    • Coding“doubling frontier-task generation across math, code, and SWE.” source ↗
    • Math“doubling frontier-task generation across math, code, and SWE.” source ↗
    • Security“a procedural generator of capture-the-flag environments for training and evaluating Unix competence in language-model shell agents.” source ↗
    • Games“GlyphBench renders over 360 game tasks as Unicode grids for language-model RL” source ↗

    What it sells

    • RL environments“It transforms your proprietary data and evals into new sets of environments, and refines agents on new examples of the tasks they are intended to perform.” source ↗

Coding 13

Writing, reviewing and running software: repositories, terminals, tests.

  • Bespoke Labs RL envsData

    Bespoke Labs builds RL environments and infrastructure, agent evaluation and optimization tools, and open datasets and benchmarks for frontier labs and enterprises training long-horizon agents.

    Why it's in Coding

    • Coding leads with“Company-scale simulations with real codebases and microservices” source ↗
    • Enterprise work“Evaluate agents in environments that mirror your systems and processes.” source ↗
    • Math“The open reasoning dataset has been downloaded hundreds of thousands of times” source ↗

    What it sells

    • RL environments“Bespoke Labs builds RL environments and infrastructure to let frontier labs and enterprises train long-horizon agents for production.” source ↗
    • Evals“Evaluate agents in environments that mirror your systems and processes.” source ↗
    • Datasets“One of the best open reasoning datasets.” source ↗
    • Human data“Domain experts interested in building RL environments and data pipelines.” source ↗
  • Datacurve RL envsData

    Datacurve sells expert-sourced training data for frontier AI labs, including RL environments, long-horizon tasks, agent trajectories, SFT demonstrations, off-the-shelf datasets and benchmarks, with a coding focus.

    Why it's in Coding

    • Coding leads with“Datacurve provides the frontier coding data that powers the world's most advanced models.” source ↗
    • Security“Software EngineeringData ScienceCyber SecurityMachine LearningResearch” source ↗
    • Enterprise work“From enterprise data systems with layered schemas, to enterprise software systems with evolving business logic and real integration constraints” source ↗

    What it sells

    • RL environments“Durable environments for measuring agentic capabilities where real work happens: long horizons, natural instructions, realistic tools, domain-specific judgment, and partial progress.” source ↗
    • Datasets“Prebuilt datasets, curated for signal, reviewed for quality, and structured to drop into your training stack without translation work.” source ↗
    • Evals“We build benchmarks that capture task-faithful, domain-sensitive lift” source ↗
    • Human data“Demonstrations that set the right behavioral prior for models, captured through bespoke tooling that lets experts work naturally” source ↗
  • General Reasoning RL envsDataInfra

    General Reasoning is an AI research lab that runs OpenReward, a platform for hosting and serving RL environments for training and evaluation, plus open tooling, benchmarks like KellyBench and the BackSearch point-in-time search API.

    Why it's in Coding

    • Coding leads with“a benchmark of 89 challenging terminal-based software engineering tasks spanning compilation, cryptography, data processing, machine learning, system administration and scientific computing.” source ↗

    What it sells

    • RL environments“Connect a GitHub repository containing Harbor tasks and OpenReward will automatically detect each task directory, generate Docker images for the sandboxes, and deploy them.” source ↗
    • Evals“A benchmark for long-horizon sequential decision making.” source ↗
    • Sandboxes and compute“Connect a GitHub repository and OpenReward scans for task directories, builds Docker images for each task, and deploys them.” source ↗
    • Datasets expanding“A frozen archive of news, SEC filings, arXiv, Wikipedia and the wider web, with point-in-time search and fetch for forecasting and reproducible evals and training.” source ↗
  • Parsewave RL envsData

    Parsewave sells custom human-authored training data (datasets, reasoning traces, RL environments and private expert-graded evals) written by senior engineers and domain specialists and calibrated to a model's solve rate.

    Why it's in Coding

    • Coding leads with“Real-world engineering tasks, each with a ground-truth answer.” source ↗

    What it sells

    • Datasets“Custom datasets, reasoning traces, RL environments and evals” source ↗
    • RL environments“Runnable tasks that score the agent’s work automatically, with rubrics where a test can’t decide.” source ↗
    • Evals“Private benchmarks, graded by experts, for places where public benchmarks are saturated.” source ↗
    • Human data“Senior engineers and domain specialists write each task from scratch, under NDA and work-for-hire agreements.” source ↗
  • Revelo RL envsData

    Revelo sells expert-authored coding data, RL environments, evals and engineer staffing to AI labs, drawing on its network of 400,000+ vetted engineers.

    Why it's in Coding

    • Coding leads with“Work across the software lifecycle, from understanding and debugging an unfamiliar codebase to building features and writing tests.” source ↗
    • Games“Building games and interactive worlds in code, from rules and physics to levels, puzzles, and player input.” source ↗
    • Science“Science done through code, such as data analysis, simulation, numerical methods, and research software.” source ↗
    • Security“Cybersecurity, enterprise languages, and other work done through code, scoped with you.” source ↗
    • Finance“Finance research questions, written and checked by financial analysts, that the AI must answer by finding the data itself.” source ↗
    • Media“Scenes, 3D models, and assets made through code and 3D tools, including geometry, materials, lighting, and animation.” source ↗

    What it sells

    • RL environments“Expert-built, containerized tasks for long-horizon agent work, with verifiers, proof of solvability, and pass rates.” source ↗
    • Evals“Domain experts judge outputs against task-specific rubrics. Each eval ends with a report recommending the data and training to close the gaps.” source ↗
    • Human data“Vetted experts from our network of 400,000+ work on your platform, managed by your team or our leads.” source ↗
    • Datasets“Pre-built, expert-authored datasets and RL environments with tasks, verifiers, and trajectories, ready to license.” source ↗
  • Emulated RL envsData

    Emulated builds post-training datasets and RL environments for software engineering and AI R&D, plus benchmarks that evaluate models on production software work.

    Why it's in Coding

    • Coding leads with“building post-training datasets for software engineering and AI R&D” source ↗
    • Science“GDPx (materials science partner)” source ↗

    What it sells

    • Datasets“We provide ready-to-use datasets across software engineering, machine learning engineering, autonomous research, and model post-training.” source ↗
    • Evals“We're introducing the Software Development Automation Benchmark: 80 tasks that grade not just code generation but the ability to deploy, debug, migrate, and maintain production systems” source ↗
    • RL environments“Each environment provides the agent with a complete, runnable system and the tooling a real engineer would use.” source ↗
  • Aviro RL envsData

    Aviro builds reinforcement learning environments and post-training datasets for long-horizon coding, computer use, knowledge work and research tasks, sold to frontier labs.

    Why it's in Coding

    • Coding leads with“Designing, building, and testing complex engineering repositories.” source ↗
    • Computer use“Navigating across apps, browsers, and operating systems.” source ↗
    • Enterprise work“Synthesizing information and creating documents, spreadsheets, slides, and reports.” source ↗
    • Science“Experimenting with autoresearch, machine learning, scientific workflows, and recursive self-improvement.” source ↗

    What it sells

    • RL environments“We turn these worlds into post-training datasets and reinforcement learning environments to train the next frontier of AI.” source ↗
    • Datasets“We turn these worlds into post-training datasets and reinforcement learning environments” source ↗
  • Idler RL envs

    Idler is a frontier data lab that builds reinforcement learning environments and bespoke evaluations from real production work for frontier AI labs.

    Why it's in Coding

    • Coding leads with“Currently, we’re scaling our environment production across coding, RSI, law, cybersecurity, and enterprise use-cases.” source ↗
    • Legal“Currently, we’re scaling our environment production across coding, RSI, law, cybersecurity, and enterprise use-cases.” source ↗
    • Security“You'll lead idler's cybersecurity work: deciding which environments we build” source ↗
    • Enterprise work“Currently, we’re scaling our environment production across coding, RSI, law, cybersecurity, and enterprise use-cases.” source ↗

    What it sells

    • RL environments“Reinforcement learning environments built from real production work” source ↗
    • Evals“Frontier Coding·Economically Valuable Workflows·Bespoke Evaluations” source ↗
  • Mechanize RL envs

    Mechanize builds high-quality, long-horizon RL environments and evals for frontier coding agents, including the GBA Eval benchmark.

    Why it's in Coding

    • Coding leads with“models carry out software engineering work such as building a feature, deploying an application, or debugging an issue in an unfamiliar codebase.” source ↗

    What it sells

    • RL environments“We build environments and evals for frontier coding agents.” source ↗
    • Evals“GBA Eval is a benchmark that measures how well coding agents can write a Game Boy Advance emulator from scratch in 24 hours.” source ↗
  • Metaphi RL envsInfra

    Metaphi AI builds and delivers RL environment task sets for AI labs through its Kennet infrastructure and Taskforge harness, alongside a coding agent (Hum) and a hosted mainframe runtime (Harwell).

    Why it's in Coding

    • Coding leads with“You: a repository, a ticket, a draft” source ↗

    What it sells

    • RL environments“Kennet is Metaphi AI’s infrastructure for building and running RL environments.” source ↗
    • Sandboxes and compute expanding“Harwell, the hosted mainframe runtime.” source ↗
  • pre.dev RL envsData

    pre.dev sells a coding agent product for building and shipping software, and also sells long-horizon RL coding tasks built from licensed private codebases, delivered as runnable environments with graded verifiers and solver trajectories.

    Why it's in Coding

    • Coding leads with“Hard, verified coding tasks. From code no model has seen.” source ↗

    What it sells

    • RL environments expanding“Long-horizon RL tasks for post-training. Tasks, not repo access: each ships in a runnable environment with a graded verifier” source ↗
    • Enterprise AI services“A coding agent made for any job — it plans ahead, connects to your tools, and keeps working across devices.” source ↗
  • Proximal RL envs

    Proximal is a data research lab that builds long-horizon coding RL environments, training data and benchmarks such as FrontierSWE for frontier AI labs.

    Why it's in Coding

    • Coding leads with“Our ultra-long-horizon coding benchmark. Frontier models clear only a fraction of its tasks.” source ↗
    • Security“Our investigation into the CyberGym benchmark and reliable cyber evaluations.” source ↗

    What it sells

    • Evals“34 tasks, 13 frontier models, 20-hour budgets. Our hardest coding benchmark yet.” source ↗
    • RL environments“How we run agentic RL across Modal GPUs and our own Kubernetes sandbox fleet.” source ↗
  • Quesma RL envs

    Quesma sells an analytics tool that breaks down and optimises a team's AI coding agent spend, and has also built training environments and benchmarks for coding agents.

    Why it's in Coding

    • Coding leads with“Know where your **AI coding spend** goes” source ↗
    • Security“Can coding agents find backdoors in compiled binaries using Ghidra and radare2?” source ↗
    • Games“OpenAI cited our Baba Is Bench work while examining what makes agent evaluations succeed.” source ↗

    What it sells

    • RL environments expanding“We built training environments for difficult, multi-hour agent tasks.” source ↗
    • Evals expanding“Can coding agents find backdoors in compiled binaries using Ghidra and radare2?” source ↗

Enterprise work 5

Knowledge work in business software: CRM, ERP, email, spreadsheets, support, operations.

  • Veris AI RL envsInfra

    Veris AI sells simulation environments that recreate an enterprise's systems, data, APIs and users so AI agents can be benchmarked, tested and trained before production.

    Why it's in Enterprise work

    • Enterprise work leads with“Veris builds a high-fidelity twin of your stack and the people who use it: every API, every database, every human in the loop.” source ↗
    • Coding“Give coding agents a stateful simulation of your dependencies, so integration tests run and issues get fixed before the PR opens” source ↗
    • Finance“FinAgent Arena: one index for AI agents on bank work, AML cases and treasury questions in one simulated bank” source ↗
    • Media“VAmoS Bench · 17 voice-agent systems on the same simulated calls” source ↗

    What it sells

    • RL environments“A simulated copy of your systems, data, APIs, and users, where AI gets tested, trained, and built against before it ever touches production.” source ↗
    • Evals“Build a benchmark around your own policies, workflows, and systems, then grade every agent you buy or build against it” source ↗
    • Sandboxes and compute“Thousands of isolated environments running concurrently, with no shared staging for teams or agents to collide over.” source ↗
  • Andon Labs RL envs

    Andon Labs builds long-horizon agent evals such as Vending-Bench and runs real AI-operated businesses as safety testbeds, and is launching Pion, a platform for agents that run businesses.

    Why it's in Enterprise work

    • Enterprise work leads with“Can AI agents run vending machine businesses?” source ↗
    • Coding“Can AI models code autonomous drones?” source ↗
    • Robotics“Butter-Bench: Evaluating LLM Controlled Robots For Practical Intelligence” source ↗
    • Security“Drone-Bench: Tracking Simple Drone Surveillance Capabilities Of Frontier Models” source ↗

    What it sells

    • Evals“our evals Vending-Bench, Blueprint-Bench and Drone-Bench allow us to compare models directly in simulation.” source ↗
    • Agent safety and verification“We study and deploy frontier AI in the real world to ensure organizations run autonomously by AI are safe.” source ↗
  • Fleet AI RL envs

    Fleet AI builds high-fidelity simulated training environments (gyms) and evals that model real-world work for frontier labs, hyperscalers and enterprises training AI agents.

    Why it's in Enterprise work

    • Enterprise work leads with“training gyms that accurately model the real world and all of its work for AI.” source ↗

    What it sells

    • RL environments“Fleet is an applied research lab dedicated to the scale out of environments - diverse and high-fidelity training gyms” source ↗
    • Evals“We see the codification of human goals for agents in evals and environments as the highest leverage human activity” source ↗
  • Ooak Data RL envs

    Ooak Data turns real, anonymized company data into digital-twin RL environments with expert-level multi-step, multi-tool enterprise tasks for training and evaluating AI agents.

    Why it's in Enterprise work

    • Enterprise work leads with“We source real enterprise data, anonymize it into digital twins, and generate reinforcement learning environments with expert-level tasks.” source ↗

    What it sells

    • RL environments“We turn real company data into RL environments where AI agents learn to work in the real world.” source ↗
    • Evals“Digital twins that let you evaluate agent performance against realistic company environments before going to production.” source ↗
  • Deeptune RL envs

    Deeptune builds RL training gyms for frontier labs: simulated environments mimicking software like Slack and Salesforce where AI agents practise coding and knowledge-work tasks.

    Now part of Mercor.

    Why it's in Enterprise work

    • Enterprise work leads with“We've built 100s of gyms that mimic popular software like Slack and Salesforce.” source ↗
    • Coding“A simulation environment where AI can practice doing work (like writing code or making a spreadsheet)” source ↗

    What it sells

    • RL environments“We build training gyms for frontier research labs.” source ↗

Media 5

Audio, voice, image, video and creative work.

  • Contra Labs RL envsData

    Contra Labs sells creative-AI evaluations, benchmarks, expert human preference data, datasets and computer-use trajectories sourced from its network of creative professionals.

    Why it's in Media

    • Media leads with“AI models go head to head on creative output. Real creatives judge the results.” source ↗
    • Computer use“Computer-use trajectories from real-world projects with reasoning.” source ↗
    • Coding“helping AI tools (across video, code, image and more) understand quality the way expert creatives do.” source ↗

    What it sells

    • Human data“A frontier human data and evaluation lab making AI better for creativity.” source ↗
    • Evals“We build the eval layer for creative AI, helping AI tools (across video, code, image and more) understand quality the way expert creatives do.” source ↗
    • Datasets“Seed datasets (image, video, UI, motion). Corrective SFT data and RLHF preference pairs.” source ↗
    • RL environments“Trajectories & RL” source ↗
  • Taste Labs RL envsData

    Taste Labs builds design-focused post-training data and evaluation infrastructure for frontier labs, and sells a Brand API/MCP that lets agents extract, search and verify brand design systems.

    Why it's in Media

    • Media leads with“We’re making the unverifiable verifiable, starting with design.” source ↗

    What it sells

    • Human data“We work with the top frontier labs to give their models taste, vision and design capabilities.” source ↗
    • Evals expanding“providing the infra products for context, eval and verification, so they can output better, more beautiful, more personalized and on brand designs.” source ↗
  • Wirestock Data

    Wirestock sells licensed and custom multimodal image and video datasets for AI training, sourced from a global network of over 700K creative professionals.

    Why it's in Media

    • Media leads with“Premium multimodal data from a global network of creative professionals.” source ↗

    What it sells

    • Datasets“Images, video, illustrations, vector art, 3D assets, and more available in our vast catalog of premium datasets.” source ↗
    • Human data“Structured to your specifications, with rubrics, verifier tasks, and expert QA and review.” source ↗
  • Exabite Data

    Exabite sells off-the-shelf and custom-scraped pre-training, post-training and RL datasets, including video, documents and global web text, to frontier AI labs.

    Why it's in Media

    • Media leads with“including video, documents, and global web text beyond Common Crawl” source ↗

    What it sells

    • Datasets“Move faster with a deep off-the-shelf catalog spanning pre-training, post-training, and RL—including video, documents, and global web text beyond Common Crawl.” source ↗
  • Verita AI RL envsData

    Verita AI runs a platform of profiled PhD-level experts who produce personalized human feedback and preference data for training AI models, with a focus on aesthetics and taste.

    Why it's in Media

    • Media leads with“We care about aesthetics in all its nuance, working on how AI understands taste and how to align it with the people it serves.” source ↗
    • Coding“Developed training data for large language models by formulating coding problems that models could not resolve, and subsequently documenting the correct solutions.” source ↗

    What it sells

    • Human data“Our platform solves this by collecting open-ended information about every expert in our community and building on each profile as they work with us.” source ↗

Security 4

Cybersecurity, capture-the-flag tasks and AI red-teaming.

  • Trajectory Labs RL envsData

    Trajectory Labs provides expert red teaming, safety evaluations and RL environments to frontier AI labs to test and improve the safety of their models and agents.

    Why it's in Security

    • Security leads with“We identified 12 task-specific breaks in Gemini 4 Argon's cyber safeguards.” source ↗
    • Coding“How robust are coding agents to attacks they've never seen?” source ↗
    • Computer use“to test how well web-search agents resist prompt injection.” source ↗

    What it sells

    • Agent safety and verification“We were the only commercial red team to jailbreak Fable 5.2's cyber safeguards, succeeding 14 times.” source ↗
    • Evals“We co-developed Search-PI Complex with Meta to test how well web-search agents resist prompt injection.” source ↗
    • RL environments“Create tasks, environments, and adversarial examples that expose consequential agent failures.” source ↗
    • Human data“We connect the world's top red teamers, researchers, and engineers to mission-critical projects at frontier AI labs.” source ↗
  • ARIMLABS RL envs

    ARIMLABS builds RLVR training environments and benchmarks for long-horizon AI agents in cybersecurity, SRE, compilation and STEM, plus LLM safety benchmarking.

    Why it's in Security

    • Security leads with“Cyber — offense and defense environments” source ↗
    • Coding“Compilation — build, toolchain, and dependency repair” source ↗
    • Science“STEM — verifiable reasoning and problem solving” source ↗
    • Enterprise work“SRE — incident response simulations” source ↗

    What it sells

    • RL environments“Carefully calibrated training environments for the agents of tomorrow.” source ↗
    • Evals“We measure what frontier models can actually achieve on the bleeding edge of long-horizon, domain-specific tasks” source ↗
    • Agent safety and verification“Testing for containment failure, deception, and hidden capability in frontier models.” source ↗
  • Gray Swan AI RL envsData

    Gray Swan AI sells AI security products: Shade automated red teaming of models and agents, Cygnal runtime guardrails that block adversarial inputs and unsafe outputs, and a crowdsourced red-teaming Arena, plus pre-release safety evaluations for frontier labs.

    Why it's in Security

    • Security leads with“Shade runs attacks that work against your models, agents, and guardrails — in your deployment context.” source ↗

    What it sells

    • Agent safety and verification“Cygnal is the runtime active behavior enforcement solution trained on the attacks that reliably work against the latest frontier models.” source ↗
    • Evals“As the leading pre-release evaluator of frontier model alignment, we deliver the best available active behavior assessment and enforcement for Enterprise AI.” source ↗
    • Human data“Gray Swan tests against attacks discovered by 23,000+ adversarial researchers.” source ↗
  • Incalmo RL envs

    Incalmo builds high-fidelity, verifiable RL environments for training AI models on realistic cybersecurity tasks, partnering with frontier labs.

    Why it's in Security

    • Security leads with“create extremely high-quality cyber environments to train models to be better at cybersecurity” source ↗

    What it sells

    • RL environments“One form of this infrastructure is RL environments that task models with solving realistic security challenges.” source ↗

Computer use 3

Operating a computer like a person: browsers, desktop apps, GUIs.

  • Markov RL envsData

    Markov sells action-labelled computer-use datasets captured from domain experts and CAD/engineering RL environments with gold outputs and rubrics for training AI agents.

    Why it's in Computer use

    • Computer use leads with“The data lab for engineering and computer-use” source ↗
    • Chip design“CAD environments” source ↗
    • Games“A world model you can drive, trained on our action-labelled Forza Horizon gameplay.” source ↗

    What it sells

    • Datasets“500,000+ hours of real work, with every action labelled and narrated.” source ↗
    • RL environments“10,000+ tasks with the brief, input files, gold output and rubrics.” source ↗
    • Human data“Off-the-shelf datasets, custom collection and RL environments with gold outputs and rubrics.” source ↗
    • Evals expanding“Introducing AutoCAD Bench” source ↗
  • Toloka RL envsData

    Toloka builds RL-gym environments, MCP replicas and computer-use mockups, and runs expert human data collection to produce trajectories, graded eval signals and datasets for training and evaluating AI agents.

    Why it's in Computer use

    • Computer use leads with“Isolated, containerized browsers and interactive web applications, instrumented for DOM/screen diffs and tool/API calls.” source ↗
    • Enterprise work“Model Context Protocol replicas of enterprise tools with realistic schemas, data flows, and permission models.” source ↗

    What it sells

    • RL environments“Push your agent on context-rich simulated environments and specialized RL-gyms.” source ↗
    • Evals“Get high-fidelity trajectories and graded eval signals for training and evaluating AI agents at scale.” source ↗
    • Human data“Real expert teams executing authentic workflows with full artifact capture across version control, project management, and communication tools.” source ↗
    • Datasets“Receive versioned datasets, eval reports, and structured outputs ready for training and benchmarking.” source ↗
  • Plato RL envs

    Plato builds simulated software 'worlds' replicating the tools and workflows of real jobs, sold to AI labs as RL environments with automatic scoring for training and evaluating computer-use agents, plus published benchmarks.

    Why it's in Computer use

    • Computer use leads with“the agent operates the VM itself: pixels, keyboard, shell, files.” source ↗
    • Enterprise work“A **sim** (simulator) is a pre-built VM image of an application — a CRM, a Git server, a wiki, a Linux desktop.” source ↗
    • Finance“Finance & Accounting: 75 worlds” source ↗
    • Legal“Legal & Compliance: 42 worlds” source ↗
    • Health and bio“Healthcare Administration: 28 worlds” source ↗
    • Coding“Rebuild a working web app by studying a running copy, on 9 reference apps.” source ↗

    What it sells

    • RL environments“Each of our worlds is a working simulation of the tools, software and systems a job runs on, built around the exact tasks and workflows people do in it.” source ↗
    • Evals“Train and evaluate your models across the worlds we've built and the work inside them.” source ↗

Finance 3

Banking, trading, accounting, insurance and financial analysis.

  • Dissei RL envsData

    Dissei sells finance RL environments, sealed evaluations and graders, agent trajectories and training data built from its own lending decisions and public records.

    Why it's in Finance

    • Finance leads with“Worked examples and training records built from real financial decisions, reviewed by practitioners.” source ↗

    What it sells

    • RL environments“A financial case as it stood on one date, with its task, tools, dataroom filesystem and verifiers, ready to run inside your infrastructure.” source ↗
    • Evals“Assess financial judgment against fixed criteria, with evidence behind each score.” source ↗
    • Datasets“Off-the-shelf reasoning data: worked examples, training records and financial alignment data from real financial decisions, reviewed by our credit team.” source ↗
  • EdotEnv RL envsData

    EdotEnv turns financial market data into datasets and research environments/benchmarks for training and evaluating AI agents on quantitative finance research.

    Why it's in Finance

    • Finance leads with“Our mission is to use financial markets at scale to build software for increasingly challenging financial time-series data.” source ↗

    What it sells

    • Datasets“Turning financial markets into datasets for enterprises” source ↗
    • Evals expanding“Alpha Autoresearch evaluates the ability of agents to generate predictive signals (alpha) from noisy market data.” source ↗
    • RL environments expanding“Markets provide realistic environments that grow more challenging over time.” source ↗
  • Halluminate RL envs

    Halluminate builds RL environments and benchmarks for AI agents doing knowledge work, starting with financial services such as due diligence and private-equity deal analysis, plus computer-use and browser-agent environments.

    Why it's in Finance

    • Finance leads with“We build the environments and benchmarks that teach AI agents to do real financial work” source ↗
    • Computer use“AI sandbox environments for training and evaluating computer-use agents on realistic simulated websites.” source ↗
    • Enterprise work“Our first area of focus is economically valuable workflows in financial services and consulting” source ↗

    What it sells

    • RL environments“Halluminate is a data research lab building benchmarks and RL environments to push the frontier of AI in knowledge work.” source ↗
    • Evals“We present Westworld Finance Diligence Bench, a benchmark of 88 problems that evaluates AI agents on a complete company acquisition due-diligence process.” source ↗

Robotics 3

Embodied and physical AI: robots, factories, manufacturing.

  • Normal RL envsData

    Normal is an applied research lab selling data and evals (benchmarks, datasets) for AI in hardware and CAD design.

    Why it's in Robotics

    • Robotics leads with“CAD Arena tests how agents turn drawings into editable parts across the tools used to build the manufactured world.” source ↗
    • Computer use“CAD Arena tests how agents turn drawings into editable parts across the tools used to build the manufactured world.” source ↗

    What it sells

    • Evals“We advance frontier AI capabilities for thosebuildingaircraftrockets, robots, aircraft, satellites, medical devices, and semiconductorswith data and evals.” source ↗
    • Datasets“Benchmarks, datasets and technical reports on machine generated CAD.” source ↗
  • Uber AI RL envsData

    Uber AI Solutions sells human data services—data collection, annotation, testing and evaluation—through a global network of over 10 million contributors and experts, for generative, agentic and physical AI.

    Why it's in Robotics

    • Robotics leads with“Build robotic and autonomous systems to operate in the real world.” source ↗

    What it sells

    • Human data“Tap into our global network of over 10 million contributors and specialized experts spanning a range of industries, languages, and disciplines” source ↗
    • Evals“accessing the best of Uber’s data collection, annotation, testing, human-in-the-loop validation, and evaluation capabilities.” source ↗
  • AIChamp RL envs

    AIChamp sells fixed-price robotics automation reviews for factories, building 3D simulations of a manufacturing station from phone video to test robot options and payback before purchase.

    Why it's in Robotics

    • Robotics leads with“Sim-to-real robotics for real factory tasks” source ↗

    What it sells

    • Enterprise AI services“Send a video of the job. In 2 weeks we test people, programmed robots and robots with brains in a 3D sim of your station” source ↗

Science 3

Physical sciences and research workflows outside biology.

  • Hillclimb RL envsData

    Hillclimb builds and sells RL environments and aggregated human research data to frontier labs to improve models' AI research capabilities.

    Why it's in Science

    • Science leads with“If you'd like your models to be better at AI research, email us at” source ↗
    • Math“we deployed a similar approach for math, fostering a dedicated community of IMO medalists” source ↗

    What it sells

    • RL environments“To further scale these ambitions, we sell the environments we make to frontier labs.” source ↗
    • Human data“Aggregating all human research data” source ↗
  • Sepal AI RL envsData

    Sepal AI built expert-written evaluation benchmarks and training data for frontier AI models using a network of PhD-level domain experts; it was acquired by Mercor.

    Now part of Mercor.

    Why it's in Science

    • Science leads with“design and review high-quality physics problem sets that test AI’s capabilities in solving real-world physics challenges” source ↗
    • Health and bio“We are developing a groundbreaking benchmark to rigorously evaluate the ability of AI systems to reason about complex biological systems and phenomena.” source ↗

    What it sells

    • Evals“We're developing novel methods to evaluate how well AI systems can reason through advanced physics workflows.” source ↗
    • Human data“Frontier AI needs frontier experts. Work with Sepal AI to fuel safer and smarter products and unlock new possibilities for humanity.” source ↗
  • dmodel RL envs

    d_model is an AI research lab that builds RL environments for open-ended interpretability and alignment research tasks, partnering with frontier labs to train agents as interpretability researchers.

    Why it's in Science

    • Science leads with“partnering with frontier labs to turn their models into capable interpretability and alignment researchers” source ↗

    What it sells

    • RL environments“We build RL environments for open-ended interpretability tasks. These environments teach agents how to do cutting-edge research, not how to reward-hack.” source ↗

Health and bio 2

Medicine, clinical work, biology and drug discovery.

  • Tacit Labs RL envs

    Tacit Labs builds data, long-horizon evaluations and RL research environments for training frontier models on biology, chemistry and drug discovery tasks.

    Why it's in Health and bio

    • Health and bio leads with“help AI systems make progress on real, long-horizon problems in biology, chemistry, and drug development.” source ↗
    • Science“Tacit Labs is building data, evaluations, and research environments unlocking scientific discovery capabilities for frontier models.” source ↗

    What it sells

    • Evals“We are starting with long-horizon evaluations across the drug discovery and development pipeline to understand exactly when and where agentic systems fail.” source ↗
    • RL environments“help build Tacit's scientific agent environments: integrated environments where language-model agents can use tools, inspect data, make decisions” source ↗
  • Latch RL envs

    Latch builds benchmarks and evals for AI agents in biology, such as TxBench, scBench and BioSecBench, alongside a bioinformatics software platform.

    Why it's in Health and bio

    • Health and bio leads with“Researching and ImprovingAI Agents for Biology” source ↗
    • Security“Biosecurity at the frontier” source ↗

    What it sells

    • Evals“We introduce TxBench–Antibody Development, a verifiable benchmark of 60 evaluations spanning the seven stages from lead nomination to the first-in-human (FIH) dose.” source ↗

Games 1

Video games and game-like simulations.

  • Good Start Labs RL envsData

    Good Start Labs turns games into RL training environments, human gameplay data, and benchmarks for AI labs.

    Why it's in Games

    • Games leads with“Games are compact worlds of skill and physics. We turn them into training materials that make models more capable.” source ↗
    • Enterprise work“training Qwen 3 on our Diplomacy environment improved it on a customer-support benchmark (Tau2)” source ↗

    What it sells

    • RL environments“We build interactive learning environments from curated games with unique, verifiable rewards, so models get measurably better at hard, useful work.” source ↗
    • Datasets“Text and image data from real people playing, captured at the source with clear provenance.” source ↗
    • Evals“Public, transparent measurement of model capability, so every gain is real and reproducible.” source ↗

Math 1

Mathematical reasoning and proofs.

  • Ulam RL envsData

    Ulam sells ultra-hard RL tasks, environments, reasoning datasets and private benchmarks in mathematics, STEM and cybersecurity to frontier AI labs.

    Why it's in Math

    • Math leads with“Difficult mathematical problems with executable verifiers and precise reward signals.” source ↗
    • Science“Scientific reasoning tasks that demand more than a plausible explanation: quantitative predictions, computational results, and testable solutions” source ↗
    • Security“Challenging security environments for training and evaluating agents on investigation, defensive analysis, and tool use” source ↗

    What it sells

    • RL environments“We help teams hillclimb Terminal-Bench Science, GDP.pdf, and other popular benchmarks with targeted RL environments.” source ↗
    • Evals“ErdosBench, STEMBench, and CyberBench test models on difficult problems beyond standard benchmarks.” source ↗
    • Datasets“100,000+ internally created, unpublished private math problems” source ↗

Infrastructure 11

Sells runtime or tooling rather than tasks or data, so it has no domain.

What's changed

  1. Vmax added Games
  2. Toloka now leads with Computer use
  3. Trajectory Labs added Computer use
  4. Modal Training platforms is now expanding
  5. Vals AI added Coding
  6. Vals AI added Legal
  7. Vals AI added Science
  8. Parsewave added Datasets
  9. Parsewave added RL environments
  10. Patronus AI no longer shows Evals
  11. Patronus AI no longer shows Media
  12. Quesma added Games
  13. Prime Intellect added Chip design
  14. Prime Intellect added Media
  15. Rise Data Labs added RL environments
  16. Rise Data Labs added Datasets
  17. Rise Data Labs added Computer use
  18. Rise Data Labs added Health and bio
  19. Rise Data Labs added Legal
  20. Rise Data Labs added Finance
  21. Rise Data Labs added Coding
  22. Refresh added Human data
  23. Scale AI added Agent safety and verification
  24. Scale AI added Security
  25. Scale AI added Media
  26. Scale AI no longer shows Finance
  27. pre.dev added Enterprise AI services
  28. Symbal now leads with no single domain
  29. Ulam added Security
  30. Turing RL environments is now expanding
  31. Turing added Security
  32. Veris AI Sandboxes and compute is now core
  33. Metaphi added Sandboxes and compute (expanding)
  34. Metaphi now leads with Coding
  35. Sepal AI added Evals
  36. Sepal AI added Human data
  37. Sepal AI added Science
  38. Sepal AI added Health and bio
  39. Sepal AI now leads with Science
  40. Normal added Computer use

Not placed yet (4): Ionic, Matrices, Originator, Phinity. Their websites don't say enough about what they sell; they are re-read every month.

How this works: each vendor's homepage and one product page are read and sorted into fixed categories, with a quote from the vendor's own site for every claim. A vendor that covers three or more domains with none leading is Multi-domain. Definitions · About the data