What are the best training data companies in 2026?
Last updated: November 9, 2026
The best training data companies in 2026 are Surge AI for top-quality expert RLHF and RL-environment data, Mercor for professional-domain experts and evaluations, and Scale AI for the widest coverage, from post-training to robotics. For specific needs, Turing leads on coding and STEM data, Snorkel AI on hard benchmarks and environments, Prolific on self-serve human evaluation with public pricing, and TELUS Digital and Appen on very large multilingual programs.
This is an independent research guide. We did not test these services ourselves. Rankings come from each company's own published materials and from reputable reporting. See our methodology.
The ranked list
1. Surge AI
What it is: A human-data company for frontier AI. Its products include agentic RL environments and verifiers, rubric design, RLHF preference data, supervised fine-tuning (SFT) demonstrations, human evaluation, expert-domain data, multilingual data in 70+ languages, multimodal image, audio and video work, and off-the-shelf datasets.
Best for: Frontier and large model labs that need very high-quality expert data and agent-training environments.
Strengths: Forbes (September 2025) reported about $1.2 billion in 2024 revenue, a business that grew without outside funding, and customers including Google, Meta, Microsoft, Anthropic and Mistral. It has moved early into RL environments, verifiers and public benchmarks.
Drawbacks: It discloses very little publicly. The homepage is mostly a mission statement, and there is no public pricing or customer directory. Its funding talks were reported but not confirmed as closed in the sources we reviewed. It is built for large, bespoke programs, not small teams.
Pricing: Custom, by quote.
Official site: surgehq.ai
2. Mercor
What it is: An expert network and research company. It builds human data, evaluations, benchmarks (the APEX family) and RL environments based on real professional work. It also offers business data monetization and enterprise AI deployment.
Best for: Labs and enterprises that need doctors, lawyers, bankers, accountants or engineers to write, grade and evaluate model work.
Strengths: Mercor says it has 5M+ experts and a $90/hr average contracted rate. It publishes partner benchmarks (Harvey LAB, APEX-Accounting, APEX-SWE, τ³-Banking) with public leaderboards. It started a research team in October 2026 that also does post-training. CNBC reported a $10 billion valuation after a $350 million round (October 2025).
Drawbacks: Expert hours are expensive, so it is a poor fit for high-volume, low-complexity labeling. Its scope is broadening across data, evals, enterprise AI and data monetization, so buyers should confirm which team will run their project.
Pricing: Custom engagements, by quote.
Official site: mercor.com
3. Scale AI
What it is: A long-established data company (founded 2016). Its Data Engine supplies post-training and evaluation data. It also runs Scale Labs leaderboards and offers physical-AI/robotics data, automotive data, public-sector work and enterprise AI applications.
Best for: Organizations that want one vendor across LLM data, evaluations, autonomy and robotics data, or government work.
Strengths: It covers the most ground of any company here. Scale says 25% of its contributors have advanced degrees and that it powers 90% of leading generative AI model builders (a company claim). Named customers include Meta, Physical Intelligence, Universal Robots, the Center for AI Safety and the U.S. DoD CDAO.
Drawbacks: In June 2025 Meta invested $14.3 billion for a 49% stake, and founder Alexandr Wang moved to Meta (AP). Labs that compete with Meta may care about vendor neutrality. Pricing is not public.
Pricing: Custom, by demo and quote.
Official site: scale.com
4. Turing
What it is: A frontier-AI data and services company. It sells custom and off-the-shelf datasets, RL environments, expert contributors and benchmarks in software engineering, enterprise knowledge work and frontier STEM, plus enterprise AI services.
Best for: Coding, agentic software-engineering and STEM reasoning data.
Strengths: It is strongest in coding, a result of its origins as a developer talent network. It shows Anthropic, Google Gemini and Nvidia among its frontier partners. It raised a $111M Series E at a $2.2B valuation (2025) and published new research in 2026, including Behavior2Code and SciCode++.
Drawbacks: It is less focused on multimodal, speech or computer-vision labeling. Its homepage gives few specifics, so ask for dataset samples.
Pricing: Custom. Dataset samples are available on request.
Official site: turing.com
5. Snorkel AI
What it is: A company spun out of Stanford AI Lab. It has moved from programmatic-labeling software to expert data development: specialized training data, evaluation environments and verifiable reward signals for agentic coding, computer use, science and knowledge work.
Best for: Teams whose models fail on hard, specialized tasks where it is difficult to define what a correct answer looks like.
Strengths: It is heavily research-driven, with published work on RLVR (reinforcement learning with verifiable rewards) and rubric failure modes. It runs many public leaderboards (Agents' Last Exam, OSWorld 2.0, Terminal-Bench 4.0, Senior SWE-bench) and has expanded its Open Benchmarks Grants tenfold.
Drawbacks: It focuses on difficulty, not volume, so it is not the choice for bulk annotation. Pricing is not public.
Pricing: Custom. Dataset samples are available on request.
Official site: snorkel.ai
6. Invisible Technologies
What it is: An AI training and evaluation vendor. It offers expert annotation, RLHF, agentic evaluation, RL environments, multimodal data and red-teaming through its Meridial expert network. It also does enterprise AI and automation work.
Best for: Labs and model builders that need PhD-level or rare-language experts quickly.
Strengths: Invisible says it has trained foundation models for more than 80% of leading AI providers (a company claim). It covers 80+ languages and publishes a case study on evaluating Cohere's Command A.
Drawbacks: Much of its business is enterprise consulting and automation, so confirm that a dedicated data-lab team will staff your project. Pricing is not public.
Pricing: Custom, by demo.
Official site: invisibletech.ai
7. Labelbox
What it is: Originally a labeling platform. Its products now include Horizon (RL environments and evals for post-training), Terra (robotics data: egocentric and teleoperated demonstrations), the Alignerr expert network and Recursion (background engineering agents).
Best for: Teams that want software plus an expert workforce from the same vendor, including robotics demonstration data.
Strengths: Labelbox says Alignerr has 2.6M+ skill-tested experts across 200+ domains. Published customer stories include Meta (the GIM benchmark), Google Cloud Vertex AI evaluation and Intuitive Surgical. It recently announced a $100M+ Series E.
Drawbacks: Its public pricing page has been removed and buying now goes through sales. Its new agent product makes the site harder to navigate if you only want labeling.
Pricing: Sales-scoped, custom.
Official site: labelbox.com
8. Prolific
What it is: A verified-participant platform plus managed services for AI human data: RLHF, preference ranking, SFT, factuality checks, domain-expert evaluation and research studies.
Best for: Researchers and AI teams that want fast, self-serve human evaluation with clear pricing.
Strengths: It has 300,000+ active participants and 300+ audience filters. Its Protocol system runs 40+ identity and behavior checks. It offers API and no-code integrations. It is the only company here with fully public pricing.
Drawbacks: In self-serve mode you design and run quality control yourself. Its expert pool is smaller than the large expert-network vendors.
Pricing: Pay-as-you-go. You pay participant rewards plus a platform fee, usually 42.8% for corporate customers and 33.3% for academic or non-profit customers. Managed services are priced by SOW. See our pricing comparison.
Official site: prolific.com
9. Toloka
What it is: A training-data company for AI agents and LLMs. It offers RL environments and simulated virtual companies, agent training datasets, evaluation and red-teaming, coding data, physical-AI and egocentric video data, off-the-shelf datasets and Toloka Arena. Its expert workforce works through Mindrift.
Best for: Agent builders that need realistic environments plus a platform that automates pipeline setup.
Strengths: Its platform builds multi-stage collection, annotation and QA pipelines from a project description. It covers many agent types (computer use, deep research, corporate assistants, coding) and publishes case studies such as poolside.
Drawbacks: It is smaller than the top-tier vendors by reported revenue. Toloka publishes its own client testimonials anonymously, so ask for referenceable customers.
Pricing: A self-service platform agreement plus quoted managed projects.
Official site: toloka.ai
10. TELUS Digital
What it is: The AI data business of TELUS Digital. It covers GenAI post-training, automotive sensor-fusion data, physical-AI/robotics collection, search and ads rating, and off-the-shelf datasets.
Best for: Large enterprises that need multilingual, multi-country or in-field data collection with enterprise compliance.
Strengths: TELUS Digital reports 1M+ contributors, 500+ annotation languages and dialects, field operations in 50+ countries, more than two billion labels a year and 20+ years of experience. It calls itself an independent, neutral partner.
Drawbacks: Large-vendor processes can be slow for small or experimental projects, and pricing is not public.
Pricing: Custom, through sales.
Official site: telusdigital.com
11. Appen
What it is: A human-data company dating to 1996. It now sells frontier-model alignment products (expert RLHF, SFT demonstrations, red teaming, agentic RL environments, tool-use traces, multimodal data) and an off-the-shelf data catalog. Its workforce works through CrowdGen.
Best for: Multilingual speech, search-relevance and large localized programs, and buyers who want ready-made datasets.
Strengths: It has 30 years of history across speech, search, translation and RLHF. It reports coverage of 80+ languages and 500 locales, and states that it measures inter-annotator agreement and documents data provenance.
Drawbacks: Its frontier-expert positioning is newer than its crowd-scale heritage, so ask for references on expert-level work.
Pricing: Custom. Catalog datasets are licensed.
Official site: appen.com
12. iMerit
What it is: An expert data company, now owned by EXL. It runs the iMerit Scholars expert network and the Ango Hub annotation and evaluation platform, covering generative AI, mobility, healthcare and robotics.
Best for: Medical imaging, LiDAR/3D sensor fusion and regulated, audit-ready data.
Strengths: iMerit reports 25,000+ domain experts in 60+ countries. Ango Hub supports DICOM, point clouds, long video and chain-of-thought workflows, and iMerit relies on a dedicated in-house workforce rather than a crowd.
Drawbacks: The EXL integration is recent. It is less visible in frontier-lab LLM post-training than the top six.
Pricing: Custom, by demo.
Official site: imerit.ai
13. Sama
What it is: A data annotation, validation and model-evaluation company founded in 2008. It works mainly on computer vision, sensor and multimodal data, and offers the Sama Annotate and Sama Curate tools.
Best for: Production computer vision in retail, ADAS, agriculture and robotics, and buyers who prioritize impact sourcing.
Strengths: Sama claims a 99% first-batch acceptance rate and an average customer tenure of 8 years. It is a certified B Corp, and its case studies include Walmart (2.5M+ SKUs).
Drawbacks: It focuses less on frontier LLM reasoning and RL-environment work.
Pricing: Custom. Timelines and pricing are scoped in a consultation.
Official site: sama.com
Comparison table
| # | Company | Best for | Core data types | Public pricing? |
|---|---|---|---|---|
| 1 | Surge AI | Frontier labs | RLHF, SFT, RL environments, 70+ languages, multimodal | No |
| 2 | Mercor | Professional-domain experts | Expert data, evals, APEX benchmarks, RL environments | No |
| 3 | Scale AI | Broad coverage, robotics, public sector | Post-training, evals, physical AI, automotive | No |
| 4 | Turing | Coding and STEM | SWE data, RL environments, datasets | No |
| 5 | Snorkel AI | Hard specialized tasks | Expert data, environments, benchmarks | No |
| 6 | Invisible | Fast expert mobilization | RLHF, evals, red-teaming, 80+ languages | No |
| 7 | Labelbox | Platform plus experts | RL environments, evals, robotics demos | No (sales-led) |
| 8 | Prolific | Self-serve human evaluation | Preferences, evals, SFT, studies | Yes |
| 9 | Toloka | Agent environments | RL gyms, agent data, red-teaming, egocentric video | Partly |
| 10 | TELUS Digital | Global enterprise scale | GenAI, automotive, physical AI, 500+ languages | No |
| 11 | Appen | Multilingual and catalog data | RLHF, speech, search, OTS datasets | No |
| 12 | iMerit | Medical and LiDAR | CV, 3D, DICOM, GenAI evals | No |
| 13 | Sama | Production computer vision | Image, video, sensor annotation, validation | No |
How to choose a training data company
- Start with the failure, not the vendor. Write down where your model breaks: reasoning, coding, a professional domain, a language, perception, or agent tool use. Match that to the vendor's specialty using the table above.
- Expert data or volume labeling. Post-training and RL environments need experts (Surge, Mercor, Turing, Snorkel, Invisible). High-volume perception labeling suits Sama, iMerit, TELUS Digital and Appen.
- Ask for a paid pilot and samples. Most of these vendors offer dataset samples. Judge them on your own rubric before you commit.
- Check quality controls. Ask about expert vetting, inter-annotator agreement, gold sets, audits and how disputes are resolved.
- Check neutrality and confidentiality. Ask who owns or invests in the vendor, whether your competitors use it, and how your data is kept separate.
- Security and compliance. Confirm data residency, certifications and whether the data you receive is licensed or owned outright.
- Budget model. Use self-serve platforms (Prolific) for small studies and managed SOWs for large programs. See pricing.
FAQ
What is a training data company?
A training data company supplies the data used to train and evaluate AI models. That includes human-labeled examples, expert demonstrations, preference rankings for RLHF, evaluation sets and RL environments. Most combine a vetted human workforce with software and quality control.
Which company is the best for LLM training data?
For frontier LLM post-training, Surge AI, Mercor and Scale AI are the strongest-evidenced choices in 2026. Turing is a strong pick for coding, and Snorkel AI for hard, specialized evaluations and environments.
Who are Scale AI's main competitors?
Surge AI, Mercor, Turing, Invisible Technologies, Snorkel AI and Labelbox compete for frontier post-training work. TELUS Digital, Appen and Toloka compete on scaled and multilingual programs.
How much does AI training data cost?
Most vendors price by custom quote. Prolific is a public exception: you pay participant rewards plus a platform fee of usually 42.8% for corporate customers. Expert work costs far more per hour than general labeling. Mercor, for example, reports a $90/hr average contracted rate for its experts.
What is RLHF data?
Reinforcement learning from human feedback (RLHF) data consists of human judgments, usually rankings or ratings of model responses. It is used to train a reward model that steers the model toward preferred behavior.
What are RL environments and why do data companies sell them?
RL environments are simulated tasks with checkable outcomes, such as a codebase, a browser or a mock company's systems. AI agents practice in them and get graded. Surge AI, Mercor, Turing, Snorkel AI, Labelbox, Toloka, Invisible and Appen all now offer them, because agent training needs them.
Is Scale AI still independent after Meta's investment?
Meta invested $14.3 billion for a 49% stake in June 2025, and founder Alexandr Wang joined Meta (AP). Scale still operates as its own company, but some buyers factor that relationship into vendor-neutrality decisions.
Which training data companies are best for computer vision?
For perception and sensor data, look at Sama, iMerit, TELUS Digital and Scale AI. For robotics demonstration data, also consider Labelbox (Terra) and Toloka.
Can I buy off-the-shelf training datasets?
Yes. Appen, TELUS Digital, Toloka, Surge AI, Mercor and Turing all offer ready-made datasets as well as custom collection.
Related guides: Best AI training data providers · Best AI datasets · Best robotics data companies