RLHF vendors compared

Last verified November 7, 2026

Most RLHF vendors sell a managed service and quote after scoping; few publish prices. The table lists what each vendor says it supplies, how it delivers (managed, platform or marketplace) and any pricing on its own pages.

RLHF and human feedback vendors, from their own pages, read November 7, 2026
Vendor	What it says it supplies	Delivery	Published pricing	Source (seen Nov 7, 2026)
Surge AI	Human data for frontier labs: preference rankings for RLHF, SFT writing, rubric and verifier design, human evals, RL environments and prebuilt expert datasets. Claims coverage of 70+ languages.	Managed service	Not published	Surge AI: Products
Toloka	Amsterdam-based. Lists preference pairs for RLHF and DPO, SFT demonstrations, red teaming, and tasks with automatic checks for evaluation and RL.	Managed service	Not published	Toloka: About
Mercor	Recruits domain experts for lab projects: human evals, expert-written training data, RL environments and its own APEX benchmark suite.	Managed service	Not published	mercor.com
micro1	Expert human data for AI labs. TechCrunch called it a Scale AI competitor and reported its claim of passing $100M in annual recurring revenue.	Managed service	Not published	TechCrunch (Dec 4, 2025)
Turing	Post-training data and model evaluation for labs. Its RL environment page shows working copies of business tools such as Jira, Salesforce and Zendesk.	Managed service	Not published	Turing: RL environments
TELUS Digital	Red teaming, generative AI post-training data and packaged datasets, from a contributor community it puts at over 1 million people in 500+ languages.	Managed service	Not published	TELUS Digital: Data for AI training
Handshake AI	Draws on the Handshake careers network: students and experts join its AI Fellowship and are paid to train and test lab models.	Managed service	Nothing published for buyers; fellows are told pay differs from project to project.	Handshake support: MOVE Program
Appen	Listed on the ASX. Annotation and post-training data; a 2026 blog describes 500 coding tasks that each come with an automated grader.	Managed service	Not published	Appen blog: RLVR (updated Oct 2026)
Innodata	Nasdaq: INOD. Labels AI datasets, tunes models for specific jobs and runs safety tests.	Managed service	Not published	innodata.com
Scale AI	Sells collection, curation and labeling of training data through its data engine, and runs model evaluations. Separately lists ready-made tool-use and computer-use RL environments. Raters reach it through Outlier, its contributor brand.	Platform plus services	Not published	Scale AI: Data Engine; Scale AI: RL Environments; Outlier: About us
Labelbox	Labeling software, with its own labeling workforce available as an extra. Its expert network goes by the name Alignerr.	Platform plus services	Metered in Labelbox Units: 20 per data row on Live Multi Modal or LLM projects, and 500 free each month on free accounts. No dollar value per unit is listed.	Labelbox docs: Billing; Alignerr: About
Kili Technology	Software for annotation, evaluation and post-training datasets; specialist labelers can be added.	Platform plus services	A $0 trial (2 to 5 seats, 100-asset cap); paid Grow and Enterprise plans are quoted.	Kili: Pricing
SuperAnnotate	Tooling to create labeled, evaluation and fine-tuning datasets.	Platform plus services	Three tiers (Starter, Pro, Enterprise), all without listed prices.	SuperAnnotate: Pricing
Centific	Its AI Data Foundry pairs a platform with services: collection, preference data for RLHF, SFT, RL environments, safety evals and a data marketplace.	Platform plus services	Not published	Centific: AI Data Foundry
Prolific	A participant pool for academic studies and AI human data. You build the study, set the pay and pay participants yourself; Prolific adds its fee.	Self-serve marketplace	Participants: $8 an hour floor, $12 suggested. Prolific's own fee is extra and its rate is not given in that article.	Prolific help center: payment model (Jul 17, 2026)

A vendor in this table can pay RLHF Companies when an introduction turns into a contract. Who appears, the grouping by delivery model and every description are not for sale.

Independent site: none of these vendors owns, funds or has approved this page. Method: each vendor's public pages were read on November 7, 2026 and only their own statements were kept; no service was tested.

How should you read this table?

"What it says it supplies" is each vendor's own description, shortened. It tells you who claims to do preference data, demonstrations or red teaming; it does not tell you who does them well. Treat it as a shortlist to scope, not a ranking.

"Not published" is common and not a warning sign. Human feedback is priced on time per task and rater expertise, so a managed vendor cannot post a rate card that means much. The published figures come from platforms and marketplaces, where you supply some of the work yourself. The pricing models page explains the structures.

What are the three kinds of RLHF vendor?

Managed service: The vendor recruits and pays raters, runs training and quality checks, and delivers labeled data against a statement of work. You write the rubric with them and review samples. Best when you need expert raters at volume and do not want to run a workforce.
Platform plus services: Software for building labeling and evaluation tasks, with the vendor's workforce sold as an add-on. Suits teams that already have raters, or want to keep the tooling while changing who labels.
Self-serve marketplace: You design the study, set pay and recruit from the participant pool. Lowest labor cost per hour and the most work for your team: screening, gold tasks and disagreement handling are yours.

What changed in vendors' published terms?

A short, dated log of changes on the pages behind the table. We add to it each time the pages are read again.

Nov 7, 2026Baseline: every page behind the table read for the first time. Where a row says "Not published", we found no figure that day.
Oct 2026Appen revised its February write-up of coding tasks graded by automated verifiers. (Appen blog)
Jul 17, 2026Latest date on the Prolific article that sets the $8 floor and $12 suggestion for participant pay. (Prolific help center)
Jul 9, 2026Mercor said it is buying Deeptune, which builds RL environments. (Mercor blog)

Questions teams ask

Which companies offer RLHF data services?

Vendors whose own pages describe RLHF, preference or post-training data work include Scale AI, Surge AI, Toloka, Mercor, Turing, TELUS Digital, Centific, Appen and Innodata as managed services, and Labelbox, Kili Technology and SuperAnnotate as platforms with services. Prolific is a self-serve marketplace for recruiting participants.

Why do so few RLHF vendors publish prices?

Price depends on time per task, rater expertise, overlap and languages, so managed vendors quote after scoping. The figures that are public tend to come from platforms (unit or seat pricing) and marketplaces (participant pay floors).

Are Outlier and Alignerr vendors?

Not in the buyer sense. Raters sign up through them, but the companies behind them, Scale AI for Outlier and Labelbox for Alignerr, are who a lab contracts with.

Is a self-serve marketplace cheaper than a managed vendor?

The per-hour labor cost can be lower, but your team writes the task, recruits and screens raters, runs quality checks and handles disagreements. Managed vendors charge for doing that work. Compare the total cost including your own team's time.

Is this a ranking or a review?

Neither. Nobody here has tested or audited these vendors, and the table carries no scores. It restates what each vendor publishes, grouped by delivery model, with "Not published" wherever a page is silent.