Facts checked on each company's own website on 24 October 2026.
The top RLHF companies in 2026 fall into a few groups. Deepen AI Services runs a dedicated, trained, in-house rater team that keeps the same people on your model every round, rates to your guidelines with a reason for each choice, and delivers in your training format. Scale AI works with AI labs, Surge AI focuses on frontier-model data, and Invisible Technologies says it has trained over 80% of the world's leading AI models. Mercor, Labelbox's Alignerr network and iMerit bring domain experts. Appen, Toloka and Sama supply preference data alongside other LLM training data, and Prolific provides screened participants for AI evaluation.
What RLHF companies supply
- Pairwise preferences. Two responses to the same prompt, the better one chosen, often with how strongly and why.
- Rankings and ratings. Several responses put in order, or each scored against a rubric.
- Rewrites. When no response is good enough, a rater writes a corrected one, saved as the preferred response.
- Reasons and critiques. A short explanation for each choice, which makes the data auditable and easier to check.
- Prompts and demonstrations. New prompts, or ideal responses written for supervised fine-tuning.
The short answer: the right company for the work
| The work | Companies | What each offers |
|---|---|---|
| Preference data for chat, writing, summaries and instructions | Deepen AI Services, Appen, Toloka, Sama | Deepen AI Services' trained raters work to your guidelines, the same people every round, with a reason for each choice. Appen, Toloka and Sama supply preference data among other LLM training data. |
| Large post-training runs for frontier models | Deepen AI Services, Scale AI, Surge AI, Invisible Technologies | Deepen AI Services grows a dedicated team for a run, calibrating new raters on known pairs first. Scale AI works with AI labs, Surge AI focuses on frontier-model data, and Invisible runs a vetted expert network. |
| Coding, maths and specialist domains | Deepen AI Services, Mercor, Labelbox's Alignerr, iMerit | Deepen AI Services quotes specialist work in your plan and names who will rate. Mercor, Alignerr and iMerit run expert networks. |
| Ratings from trained raters or from the public | Deepen AI Services, Prolific | Deepen AI Services provides trained raters who learn your guidelines. Prolific provides screened participants from the public, self-serve. |
| Feedback in many languages | Deepen AI Services, Appen | Deepen AI Services works in English and quotes other languages. Appen cites coverage of more than 500 locales. |
The companies compared
| Company | RLHF work | How you start | Pricing |
|---|---|---|---|
| Deepen AI Services | Pairs, rankings, scores and rewrites, each with a reason, by the same trained raters every round | A plan by email in 2 working days, then a two-week pilot | Quoted in your plan before you pay |
| Scale AI | RLHF, human data, model evaluation, safety and alignment | Book an intro call | No prices on its site |
| Surge AI | Frontier data, post-training runs and RL environments | Contact its team | No prices on its site |
| Invisible Technologies | Training data and RL environments, with a network of vetted experts | Book a demo | No prices on its site |
| Mercor | Expert human data, evaluations and benchmarks | Book a demo or contact sales | No client prices on its site |
| Labelbox (Alignerr) | Experts who write rubrics, preferences and assessments | Talk to its team | No prices on its homepage |
| iMerit | RLHF, chain-of-thought, evaluation and red teaming | Talk to an expert or book a demo | No prices on its site |
| Appen | SFT pairs, preference data, prompts and reasoning traces | Talk to an expert | No prices on its site |
| Toloka | Post-training and preference data for agents and LLMs | Talk to an expert | No prices on its site |
| Sama | Preference ranking, instruction following and factuality | Talk to an expert | Discussed in a first call |
| Prolific | AI evaluation and feedback from screened participants | Self-serve | Participant pay plus 42.8% (33.3% for academics) |
The companies in detail
1. Deepen AI Services: a dedicated rater team that stays on your model
Deepen AI Services gives an AI lab or product team its own trained, in-house raters. They learn your guidelines, rate a set of known pairs before live work, and stay on your project, so your rubric and edge cases aren't relearned every round. A named team lead decides unclear items or asks you, and agreement is in a weekly report.
- Best for: Ongoing preference data and evaluations for general chat, writing, summaries and instructions, plus coding, maths and specialist domains quoted in your plan.
- What you get: Pairwise choices with their strength, rankings, scores per response, or a mix, with the reason for each choice. When no response is good enough, raters write a corrected one.
- Consistency: Raters calibrate on known pairs before live work, and a share of pairs is rated by more than one person so agreement can be checked.
- Formats: The one your training pipeline reads, such as JSON lines or CSV, with the fields you set, in your annotation tool or a shared sheet.
- Pricing: Quoted in your plan, because the rate depends on the expertise and the length of the responses. You pay for the hours raters work, with a team lead, QA checks and a weekly report included.
- How you start: Send your guidelines and 20 to 30 pairs you've already rated. You get a plan and a price by email within 2 working days, with no sales call, then a two-week paid pilot on your real prompts.
- Security: Your prompts, outputs and rated data stay under NDA, with the access you set. SOC 2 Type II, ISO 27001, ISO 9001 and TISAX.
See RLHF preference data and human data for AI labs.
2. Scale AI: the Generative AI Data Engine
Scale AI's Generative AI Data Engine covers RLHF, human data, model evaluation, safety and alignment, and Scale works with AI labs, governments and the Fortune 500. Its site lists no prices, and you start by booking an intro call.
3. Surge AI: data for frontier models
Surge AI focuses on frontier data: off-the-shelf datasets, post-training runs, RL environments and an expert workforce. Its site shows no prices, and you start through its contact page.
4. Invisible Technologies: training data with vetted experts
Invisible blends its own talent with automation to deliver training data, builds RL environments verified with domain experts, and runs Meridial, a network it describes as 24,000+ vetted experts. It says it has trained over 80% of the world's leading AI models. You start by booking a demo.
5. Mercor: experts for human data and evaluations
Mercor connects AI labs with experts, such as physicians, lawyers and engineers, for human data, evaluations and benchmarks. It publishes what it pays experts by role, averaging $100 an hour, but not what clients pay.
6. Labelbox: the Alignerr expert network
Labelbox's Alignerr is a network it describes as 2.6 million skill-tested experts in more than 200 domains, who write rubrics, preferences and assessments. Labelbox also builds RL environments and evaluations for post-training. Its homepage shows no prices.
7. iMerit: generative AI services with domain experts
iMerit, now part of EXL, offers RLHF, chain-of-thought reasoning, evaluation, red teaming and prompt and response writing, and cites more than 25,000 domain experts in its Scholars program. You start by talking to its team or booking a demo.
8. Appen: LLM training data across languages
Appen supplies custom LLM training data: SFT pairs, preference data, prompts, reasoning traces and domain expert content, with coverage of more than 500 locales. Its site shows no prices.
9. Toloka: data for AI agents and LLMs
Toloka focuses on training and evaluation data for AI agents and LLMs, including post-training and preference data, red-teaming and expert work. Its site points businesses to a conversation with its experts and shows no prices.
10. Sama: preference ranking alongside annotation
Sama's generative AI work includes text annotation for instruction following, preference ranking and factuality, next to its vision and sensor annotation. Pricing, scope and timelines are discussed in a first call.
11. Prolific: screened participants for AI evaluation
Prolific gives AI teams self-serve access to more than 300,000 active participants for evaluations and feedback, with managed services quoted separately. You pay the participants plus a platform fee, usually 42.8% for companies, and design the task and review the answers yourself.
How to choose an RLHF company
- Can the raters judge your responses? General chat needs careful, trained raters. Code, maths, medicine or law need people who know the field.
- Do the same raters stay on your project? A team that keeps your rubric between rounds gives you data that's consistent over time.
- How is agreement checked and reported? Ask how raters are calibrated, how much is rated twice, and how you'll see the results.
- Do you get reasons and rewrites? A reason for each choice makes the data auditable, and rewrites fix pairs where neither response is good.
- Will it arrive in your format? JSON lines, CSV or your own schema, in your tool or a shared location.
- Are your prompts safe? Unreleased model outputs are sensitive. Look for an NDA, in-house raters and audited controls such as SOC 2 and ISO 27001.
What to send for a pilot
- Your rating guidelines, or what makes one response better than another
- 20 to 30 pairs you've already rated, with your choices
- The format your training pipeline reads
- Where the pairs live: your annotation tool or a shared sheet
- How many comparisons you need each week, and by when