What are the best AI training data providers in 2026?

Answer in brief: For an AI lab buying human data in 2026, the best AI training data providers are Surge AI for expert RLHF and post-training data, Mercor for professional-domain experts, evaluations and RL environments, and Scale AI for the broadest full-stack data engine. Handshake AI and Labelbox are close behind. Prolific is the most transparent self-serve option for human feedback and evals on a budget. Choose by data type first (SFT demonstrations, preference data, rubrics, RL environments or evals), then run a paid pilot against your own metrics.

Last updated: November 9, 2026

Top pick: Surge AI. Surge AI supplies expert post-training data for coding, agentic tool use, professional work, instruction following and STEM reasoning, and it publishes before-and-after results from post-training open models on its own data. Its off-the-shelf datasets come with a rare guarantee: train on the full dataset first and pay only if your metrics improve. · Read the Surge AI review

This guide is written for one buyer: a data, post-training or evaluation lead at an AI lab choosing who will supply human data. That means RLHF and SFT data, expert annotation, rubrics, red teaming and evaluation. It is independent research. We did not test these services. The ranking draws on each provider's own published material and on current reporting, as explained in our methodology. Company claims, such as customer counts or network sizes, are labeled as claims. Looking for something else? For computer-vision and sensor labeling vendors see besttrainingdatacompanies.com. For robotics data see bestroboticsdataproviders.com, and for ready-made datasets see bestaidatasets.com.

The ranked list

1. Surge AI

What it is: A human-data company that supplies expert post-training data (RLHF, SFT, evaluation), off-the-shelf datasets and benchmarks. Its Training Lab post-trains open models on Surge data and publishes the before-and-after results.
Best for: Frontier and open-model labs that need high-skill human data for coding, agentic tool use, instruction following, professional work and STEM reasoning.
Standout strengths: Unusually public evidence. Surge reports that training Kimi K2.7 on 1,700 coding tasks lifted SWE-Marathon from 5.0% to 25.0%. It also reports that Nemotron 3.5 Super improved on 12 benchmarks after training on Surge data. It publishes 13 benchmark leaderboards, including DAYJOB, sudo L7, GDP.pdf, HANDBOOK.md and Riemann-bench. Its off-the-shelf data can be trained on first and paid for only if your metrics improve.
Honest drawbacks: The headline gains are self-reported. The homepage is mission copy with little operational detail, and pricing for custom projects is not published. The focus is post-training, not commodity labeling.
Pricing model: Custom quotes for bespoke work. Off-the-shelf datasets are sold on a pay-only-if-it-improves-your-metrics basis.
Official site:

2. Mercor

What it is: An expert marketplace and human-data company. It recruits professionals such as physicians, lawyers, engineers and consultants to produce training data, evaluations and RL environments, and it publishes the APEX family of benchmarks.
Best for: Labs that need data or evals built on real professional work in law, accounting, engineering, finance and software.
Standout strengths: Mercor says it has 5M+ experts and lists a $90/hr average contracted rate. Its partner benchmarks include Harvey LAB (with Harvey), APEX-SWE (built with Cognition) and τ³-Banking (from Sierra). CNBC reported a $10B valuation after a $350M round in October 2025. TechCrunch reported that labs including OpenAI, Anthropic and Meta use Mercor's experts.
Honest drawbacks: Pricing for buyers is not published. Mercor also sells enterprise AI deployment and data monetization, so confirm which team will staff your work.
Pricing model: Custom engagements, priced on request.
Official site:

3. Scale AI

What it is: The longest-established large-scale training data company, operating since 2016. Its Data Engine and Generative AI Data Engine cover data generation, RLHF, red teaming and evaluation, and it runs leaderboards with private benchmarks.
Best for: Large programs that want one vendor for many data types, plus evaluation and leaderboard work.
Standout strengths: The broadest coverage on this list. Scale says 25% of its contributors have advanced degrees and that 90% of the world's leading generative AI model builders are powered by Scale. Published partners include Meta and the Center for AI Safety.
Honest drawbacks: In June 2025 Meta invested $14.3B for a 49% stake, according to AP, and founder Alexandr Wang moved to Meta. TIME and TechCrunch reported that OpenAI and Google wound down work with Scale. OpenAI said its phase-out began before the deal. TechCrunch reported about 200 layoffs, mostly in data labeling, in July 2025. Scale says it remains independent. Labs that compete with Meta should do their own confidentiality diligence. Pricing is not public.
Pricing model: Demo-led custom contracts.
Official site:

4. Handshake AI

What it is: The AI-data arm of the Handshake careers network. It draws on students, graduates, PhDs and professionals through the Handshake AI Fellowship to produce expert data, evaluations and RL environments.
Best for: Frontier labs that need large pools of credentialed, U.S.-based experts across many academic and professional domains.
Standout strengths: Handshake says 8 of the 8 largest frontier labs train with its datasets. It also says its network holds 30M+ experts and that it employs 200+ in-house researchers and engineers. Its fellowship page lists $500M+ in payouts and 188k+ fellows. TIME reported its CEO saying demand tripled after the Meta and Scale deal.
Honest drawbacks: Fellows must be based in the U.S., which limits multilingual and global coverage. Pricing for buyers is not published.
Pricing model: Custom, via a scoping session.
Official site:

5. Labelbox

What it is: A data platform that is also a managed expert-data provider. Its products are Alignerr (an expert network), Horizon (RL environments and evals for reasoning, tool use and computer use), Terra (robotics data) and Recursion (agents for engineering and research work).
Best for: Labs that want expert preference data, rubrics and RL environments from one vendor.
Standout strengths: Labelbox says Alignerr has 2.6M+ contributors across 200+ domains, including 50K+ PhDs and 85K+ licensed professionals, and that it is SOC 2 Type II certified. Published customer stories include Meta (the GIM benchmark) and Google Cloud (human preference signal in Vertex AI). It recently announced a $100M+ Series E.
Honest drawbacks: The product line has widened fast, and the homepage now leads with an agent product. Its /pricing/ page returned 404 on 2026-11-09.
Pricing model: Sales-led; no public price list found.
Official site:

6. micro1

What it is: A self-described data lab for training frontier models and evaluating agents. Its products are Realm (RL environments), Cortex (contextual evaluation of agents in production) and real-world robotics data.
Best for: Labs that want vetted expert human data combined with agentic RL environments.
Standout strengths: An integrated expert marketplace covering AI-led vetting, QA and payments. It publishes benchmarks such as Dual-Arm Sim, MissionReport and a Medical Index. micro1 says it passed a $1B annual run rate in October 2026, and Sacra's estimate matches. Its Series B was priced at a $4B valuation.
Honest drawbacks: It is a young company. Its growth figures are run rates, not audited revenue. In August, TechCrunch put its gross run rate at $500M and estimated its net run rate at $150M–200M. Pricing is not public.
Pricing model: Custom.
Official site:

7. Turing

What it is: A company that supplies frontier labs with expert datasets, coding and reasoning traces, benchmarks and RL environments, and helps enterprises deploy agentic systems.
Best for: Coding, reasoning and agentic post-training data from software-engineering and STEM experts.
Standout strengths: Its Frontier AI pages describe human-verified coding traces for SFT and reward modeling, STEM sets with Lean-based proofs, and RL environments that replicate tools such as Jira, Salesforce and Zendesk. Dataset samples are available on request. TIME reported its CEO saying Turing added about $50M in potential contracts within two weeks of the Meta and Scale deal.
Honest drawbacks: The homepage is sparse; the detail lives on its /frontier-ai pages. It does not market a separately branded RLHF program.
Pricing model: Custom, starting with dataset samples.
Official site:

8. Snorkel AI

What it is: A Stanford AI Lab spin-out that positions itself as a frontier AI data lab. It sells ready-to-use datasets (the Snorkel Data Series), custom data development, evaluation environments and specialized agents.
Best for: Labs with hard, specialized failure modes, such as finance, insurance, legal or agentic coding, that need rubric-guided or verifiable data.
Standout strengths: Strong research output. Its leaderboard hosts 18 active benchmarks, including Agents’ Last Exam, OSWorld 2.0, Terminal-Bench 4.0, SnorkelLegal and SnorkelFinance 2.0. It publishes on RLVR and on how evaluation rubrics fail (RIFT). It says it covers 1,000+ expert-level domains and measures how well its reviewers agree with each other.
Honest drawbacks: Its human workforce is smaller than those of the marketplace leaders, and pricing is not public.
Pricing model: Custom; dataset samples on request.
Official site:

9. Invisible Technologies

What it is: An AI training and enterprise AI company. Its products are Meridial (an expert network), Synapse (evaluations), Neuron (a data platform), Atomic (a process builder) and Axon (agents), and it also builds RL environments.
Best for: Labs that want a managed partner able to stand up expert evaluation and fine-tuning projects quickly, including in many languages.
Standout strengths: Invisible says it has trained foundation models for more than 80% of the world's leading AI providers, that it has 24,000+ vetted experts, and that it covers 80+ languages. It publishes a Cohere case study on evaluating Command A.
Honest drawbacks: Much of its current marketing targets enterprise automation rather than training data. Its expert pool is smaller than those of the largest marketplaces. Pricing is not public.
Pricing model: Custom, demo-led.
Official site:

10. Toloka

What it is: A training-data provider for AI agents and LLMs. It offers RL environments, agent trajectories, RLHF and preference data, instruction tuning, coding data, evaluation and red teaming. Its expert community runs through Mindrift.
Best for: Agent builders who need realistic environments and trajectories, and labs that want a self-service platform alongside managed work.
Standout strengths: Toloka says it has 200,000+ experts across 90+ domains. Its self-service platform builds a data pipeline from a plain-language goal and checks every output with automated LLM-based QA. It publishes a self-service agreement, runs the Toloka Arena leaderboard, and has a published case study with poolside.
Honest drawbacks: Customer quotes on its homepage are anonymous, and managed pricing is not published.
Pricing model: Managed projects are quoted; a self-service platform is also offered.
Official site:

11. Prolific

What it is: A platform of verified human participants for AI evaluation, preference data and research, with an optional managed service.
Best for: Academic labs, startups and safety or eval teams that need representative human feedback quickly without a contract.
Standout strengths: The most transparent pricing on this list. Prolific reports 300,000+ active participants and 300+ audience filters, offers AI Taskers, Domain Experts and API access, and runs a quality system with 40+ checks. Ai2 is a published customer.
Honest drawbacks: The platform fee is high (usually 42.8% for corporate customers), and you design and QA the tasks yourself unless you buy managed services.
Pricing model: Pay as you go: the participant reward you set plus a platform fee (usually 42.8% corporate, 33.3% academic or non-profit). The minimum reward is $8/hr and $12/hr is recommended. Managed services are quoted under a statement of work.
Official site:

12. iMerit (an EXL company)

What it is: An expert data and annotation provider, now owned by EXL. It runs the iMerit Scholars expert network, the Ango Hub platform with its Deep Reasoning Lab, and in-house annotation teams.
Best for: Chain-of-thought reasoning data, RLHF, red teaming and evaluation in regulated domains, especially medicine.
Standout strengths: iMerit says it has 25,000+ domain experts in 60+ countries. Its case studies include a 1,000-problem chain-of-thought math corpus. Its compliance page lists SOC 2 Type 2, ISO 27001, ISO 9001:2015, HIPAA, GDPR and TISAX.
Honest drawbacks: Integrating into EXL may change account structures. Pricing is not public.
Pricing model: Custom; demo-led.
Official site:

13. Appen

What it is: A veteran human-data company whose history goes back to 1996. It offers frontier-model alignment data, an off-the-shelf catalog and the CrowdGen workforce. The alignment services are chain-of-thought traces, expert RLHF, SFT demonstrations, red teaming, rubric design and a multilingual LLM-as-a-judge service.
Best for: Multilingual RLHF, SFT and evaluation programs.
Standout strengths: 30 years of experience across speech, machine translation (100+ language pairs), search evaluation and RLHF. It publishes a case study of Cohere scaling preference-based fine-tuning: 2,400+ expert contributor hours over 12 weeks.
Honest drawbacks: Its homepage gives few specifics, and newer expert-first firms dominate recent post-training reporting. Pricing is not public.
Pricing model: Custom quotes; catalog datasets are licensed.
Official site:

14. Sama

What it is: A Certified B Corporation, operating since 2008, that provides human-verified annotation, validation and model evaluation. Its generative AI services include preference ranking, instruction-following checks, fact checking, RAG evaluation and response rewriting.
Best for: Labs that want a managed evaluation and rewrite loop with documented quality targets, and buyers for whom workforce ethics is a priority.
Standout strengths: Sama reports a 99% first-batch acceptance rate and says customers stay an average of 8 years. Published case studies include Walmart, and it recently achieved B Corp recertification.
Honest drawbacks: Its center of gravity is computer vision and sensor data rather than frontier post-training by credentialed experts. Pricing is not public.
Pricing model: Custom, scoped by project.
Official site:

Comparison table

#	Provider	Model	Best for	Notable specialties	Self-serve	Public pricing
1	Surge AI	Managed expert data + OTS datasets	LLM post-training	RLHF, coding, agentic, benchmarks	No	OTS pay-if-it-improves; else quote
2	Mercor	Expert marketplace	Professional-domain data and evals	APEX benchmarks, RL environments	No	Quote
3	Scale AI	Full-stack data engine	Many data types at scale	RLHF, red teaming, leaderboards	No	Quote
4	Handshake AI	Expert network (U.S.)	Credentialed experts at volume	Expert data, RL environments, research	No	Quote
5	Labelbox	Platform + managed experts	Preference data and agent environments	Alignerr, Horizon	Partly	Not found
6	micro1	Expert marketplace + environments	Agentic RL data	Realm, Cortex	No	Quote
7	Turing	Expert network	Coding and reasoning data	Coding traces, RL environments	No	Quote
8	Snorkel AI	Data lab	Hard specialized domains	Rubrics, verifiable data, leaderboards	No	Quote
9	Invisible	Managed data + enterprise AI	Fast-start managed projects	Meridial, Synapse, 80+ languages	No	Quote
10	Toloka	Managed + self-service	Agent environments and trajectories	RL gyms, red teaming, LLM QA	Yes	Quote (managed)
11	Prolific	Participant platform + managed	Human feedback and evals on a budget	Representative samples, preference data	Yes	Yes: reward + 42.8%/33.3% fee
12	iMerit (EXL)	Managed experts + Ango Hub	Regulated-domain reasoning data	Scholars, Deep Reasoning Lab	Platform	Quote
13	Appen	Managed + dataset catalog	Multilingual alignment data	Expert RLHF, SFT, LLM-as-a-judge	No	Quote
14	Sama	Managed annotation (B Corp)	Managed evaluation and rewrites	Validation, preference ranking	No	Quote

How to choose an AI training data provider

Name the data type first. SFT demonstrations, RLHF preference data, rubrics, agent RL environments, evaluations and red teaming are different products, and few vendors are best at all of them.
Match the expertise to the task. Frontier reasoning data needs real domain experts, such as physicians, lawyers and senior engineers. Ask how a vendor vets experts, what it pays them and who reviews their work.
Run a paid pilot against your own evals. Ask for samples (Turing and Snorkel offer them), score them against a gold set, and measure agreement between annotators and model lift. Surge's pay-if-it-improves offer on off-the-shelf data is one way to cap that risk.
Check neutrality and confidentiality. If you compete with a vendor's investor or largest customer, ask about data segregation, walls between project teams and contractual protections.
Verify security and compliance. Look for documented SOC 2 and ISO 27001 coverage, and HIPAA where clinical data is involved. iMerit and Labelbox list theirs. Review each vendor's trust center.
Fit the buying model to your scale. Startups and academic teams can begin with self-serve platforms such as Prolific or Toloka. Large programs usually run under statements of work with managed teams.
Consider workforce ethics. Ask about pay rates, rate stability and working conditions. Mercor publishes an average contracted rate, Handshake posts hourly rates on its listings, and Sama is a Certified B Corporation.

Frequently asked questions

What is an AI training data provider?

An AI training data provider is a company that supplies the human-created or human-verified data used to train and evaluate AI models. For AI labs that mainly means expert-written prompts and answers (SFT), RLHF preference rankings, rubrics, red-teaming results, agent trajectories, RL environments and human evaluations.

Who is the biggest AI training data provider in 2026?

There is no audited league table, because most of these companies are private. Scale AI was valued at $29B in Meta's 2025 investment. micro1 says it passed a $1B annual run rate in October 2026. TechCrunch reported roughly $500M in annualized revenue for Mercor in October 2025. These figures come from different periods and use different definitions.

Is Scale AI still independent after Meta's investment?

Meta bought a 49% stake for a reported $14.3B in June 2025, and founder Alexandr Wang joined Meta. Scale says it remains independent. OpenAI and Google were reported to have wound down work with Scale afterwards, so labs should do their own confidentiality diligence.

How much does AI training data cost?

Most providers quote custom prices. Prolific is the main exception: you pay the participant reward you set (at least $8/hr, with $12/hr recommended) plus a platform fee of usually 42.8% for corporate or 33.3% for academic customers. Expert work costs more; Mercor lists a $90/hr average contracted rate.

What is RLHF data and who provides it?

RLHF (reinforcement learning from human feedback) data consists of human rankings or ratings of model outputs, used to train reward models and align model behavior. Surge AI, Scale AI, Mercor, Handshake AI, Labelbox, Toloka, iMerit, Appen and Sama all offer it, and Prolific supports preference collection.

What is the difference between a labeling platform and a managed data provider?

A labeling platform is software your own team or workforce uses to create data. A managed provider supplies the experts, project management and quality assurance and delivers finished datasets. Labelbox, iMerit, Toloka and Prolific offer both.

Where can I compare robotics data or off-the-shelf datasets?

This site covers human data supplied to AI labs. For robotics and physical-AI data see bestroboticsdataproviders.com. For ready-made datasets see bestaidatasets.com.

Which provider is best for a small budget or academic research?

Prolific is the most accessible: no contract or subscription, pay as you go, and a discounted 33.3% fee for academic and non-profit customers. Toloka also offers a self-service platform under a published agreement.

How do I evaluate a training data vendor's quality?

Request samples, run a small paid pilot and score the output against your own gold set and evals. Measure agreement between annotators and check whether training on the data improves your target metrics. Also confirm how the vendor vets experts and handles security.

Sources and further reading

Each review page lists its own sources. See our methodology for how we rank, our pricing comparison for each vendor's buying model, and about for who we are. Vendors can learn about placements on get featured.

Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.