How RLHF work is priced
Last verified November 7, 2026
RLHF vendors charge in one of four ways: per task, per hour of rater time, through platform fees with labor on top, or for a fixed statement of work. Managed vendors quote after scoping; the only public figures come from platforms and marketplaces, such as Prolific's $8 hourly minimum for participants.
What pricing models do RLHF vendors use?
- Per task
A price for each comparison, ranking, scored response or written demonstration. Predictable for budgets. It only works once both sides know how long a task takes, so it usually follows a pilot.
- Per rater hour
You pay for time, at a rate set by the expertise required. Common for pilots, red teaming and expert work where task length varies widely.
- Platform plus labor
A subscription, seats or usage units for the tooling, and the workforce billed separately or brought by you. Labelbox, for example, bills usage in Labelbox Units and sells labeling services as an add-on.
- Participant pay plus fee
On a marketplace you set the pay and the platform adds a fee. You design the task and run quality control yourself.
- Fixed scope
A statement of work for a defined deliverable (say, a set number of pairs at an agreed agreement level) with milestones. Shifts the risk of slow tasks onto the vendor, who prices that risk in.
Which RLHF vendors publish prices?
Few do. These are the figures on vendors' own pages on November 7, 2026. Every managed service in our vendor table quotes on request.
| Vendor | Model | Published figure | Notes | Source |
|---|---|---|---|---|
| Prolific | Participant pay plus platform fee | $8 an hour minimum; $12 an hour recommended | Prolific adds its own fee, whose rate the article leaves out | Prolific payment model article |
| Labelbox | Usage units (LBUs), services extra | 20 units per data row (Live Multi Modal, LLM) | Free accounts receive 500 units monthly; the dollar value of a unit is not listed | Labelbox billing docs |
| Kili Technology | Platform plans, services optional | $0 trial: 2 to 5 seats, 100 assets | Grow and Enterprise tiers quoted individually | Kili Technology pricing page |
| SuperAnnotate | Platform plans | Not published | Three named tiers, none with a price beside it | SuperAnnotate pricing page |
| Handshake AI | Hourly pay to fellows | Not published | Fellows are told rates depend on the project | Handshake help article |
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.
What makes RLHF data more expensive?
- Time per task. The biggest driver. Long agent traces, code that must be run, and multi-turn conversations take many times longer than a short chat comparison.
- Who can judge. Generalists cost least. Raters who can verify code, math, law or science cost more, and specialists with advanced degrees more again.
- Overlap. Each extra rater on an item repeats the labor. Overlap a share of items for agreement figures rather than all of them.
- Review layers. Senior review of a sample, gold tasks and calibration sessions all take paid time.
- Languages. Raters in less common languages are harder to recruit and qualify.
- Exclusivity. Data made only for you costs more than data a vendor can reuse. For RL environments, interviewees told Epoch AI that exclusive deals run about four to five times the non-exclusive price (their estimates, not contract figures).
How many rater hours does a project need?
This calculator works out the labor floor: rater hours and the participant pay they imply at a marketplace rate. A managed vendor's price adds recruiting, management, review and margin on top.
Formula: items × raters per item × minutes per item ÷ 60 = hours; hours × hourly pay = rater pay. The defaults (2,000 items, 6 minutes, 20% of items sent to a second rater, $12 an hour) are an example; $12 is Prolific's recommended participant rate, read November 7, 2026. This is an estimate of labor only and not a quote: vendor prices for the same work are usually far higher, and expert raters cost several times more.
Questions teams ask
How much does RLHF data cost?
Managed vendors do not publish rates; they quote per task or per hour after seeing your task. The public floor is participant pay on self-serve marketplaces: Prolific sets a minimum of $8 an hour and recommends $12, plus its platform fee. Expert raters, overlap, QA and management all add to that.
Is per-task or per-hour pricing better for the buyer?
Per task makes budgets predictable but rewards speed; per hour rewards care but makes budgets depend on how long tasks really take. Many teams pilot on hourly terms to learn the real time per task, then agree a per-task price for production.
Why does overlap matter so much for cost?
Every extra rater on an item repeats the labor. Sending every item to three raters roughly triples the labor bill. Most projects overlap a share of items to measure agreement, rather than all of them.
Do vendors charge more for exclusive data?
Usually. In interviews Epoch AI published in January 2026, people in the RL environment market described exclusive deals as costing about four to five times the non-exclusive kind. Those are estimates from interviewees, not contract prices, and they concern environments rather than preference data.
Can you give us a price?
No. We are an introduction service and do not quote on a vendor's behalf. Send a brief and the vendor we introduce you to will price it.
Sources and dates
- Prolific help center on its payment model (article dated Jul 17, 2026), page seen Nov 7, 2026.
- Labelbox docs on billing, page seen Nov 7, 2026.
- Kili plans and pricing, page seen Nov 7, 2026.
- SuperAnnotate plans, page seen Nov 7, 2026.
- Handshake support article on the MOVE Program, page seen Nov 7, 2026.
- Denain and Barber for Epoch AI (January 12, 2026), FAQ on RL environments: exclusivity estimate from interviews.
Get a quote
Sketch the feedback work in about a minute. What comes back is one vendor name, not a round of cold calls.
RLHF Companies introduces teams to vendors, and the vendor can pay us for that introduction. Sending a brief costs your team nothing.
Keep prompts, model outputs, datasets and samples out of this form; describe the work in words instead.
Received. Here is what happens next.
- Your brief is compared against the services and terms each vendor states on its own pages.
- An email names the one vendor that looks like the best fit, with the reason.
- Reply yes and they receive your brief and contact details; reply no and nothing moves. Any pilot, quote or contract is agreed directly between your team and theirs.