How much do training data companies charge in 2026?
Last updated: November 9, 2026
Most training data companies do not publish prices. They scope each project and quote it, usually under a statement of work (SOW). Prolific is the main exception, with public pay-as-you-go pricing. The table below shows what each company publicly says about how it charges.
Pricing models by company
| Company | Public pricing model | What is published |
|---|---|---|
| Prolific | Pay-as-you-go platform; custom managed services | Participant reward plus platform fee: usually 42.8% for corporate customers, 33.3% for academic and non-profit customers. Recommended reward at least £9.00/$12.00 per hour; minimum £6.00/$8.00 per hour. No subscription required. |
| Mercor | Custom | Reports a $90/hr average contracted rate paid to experts. This is the expert's pay, not the price to buyers. |
| Labelbox | Sales-scoped | Its former public pricing page now returns 404. Its Alignerr expert site advertises expert pay of roughly $40–$120/hr. |
| Toloka | Self-service agreement plus custom projects | Links to a self-service agreement and an Arena data pricing page. Managed projects are quoted. |
| Surge AI | Custom | None |
| Scale AI | Custom (demo) | None |
| Turing | Custom; dataset samples on request | None |
| Snorkel AI | Custom; dataset samples on request | None |
| Invisible | Custom (demo) | None |
| TELUS Digital | Custom; off-the-shelf datasets licensed | None |
| Appen | Custom; catalog datasets licensed | None |
| iMerit | Custom (demo) | None |
| Sama | Custom, scoped in consultation | None |
What drives the cost
- Expertise required: Physicians, lawyers and senior engineers cost far more per hour than general annotators. Contributor pay on these platforms ranges from under $20/hr for general evaluation roles to well over $100/hr for specialist roles.
- Task complexity: Multi-turn agent trajectories, RL environments and verifiers cost more than simple classification.
- Quality overhead: Multi-stage review, gold sets and consensus labeling multiply the human hours.
- Language and locale: Rare languages and in-field collection raise costs.
- Exclusivity: Custom data you own costs more than licensing an off-the-shelf dataset.
Tips for buyers
- Ask for a pilot priced per unit, such as per accepted task, comparison or trajectory, rather than per hour.
- Define acceptance criteria and rework terms in the SOW.
- Compare off-the-shelf datasets before commissioning custom data.