Data annotation costs vary wildly depending on task complexity, geographic location, and quality requirements—and picking the wrong pricing model can blow your budget or deliver unusable labels. Understanding what you're actually paying for separates smart ML teams from those drowning in cost overruns.
Why Annotation Pricing Isn't Straightforward
Data annotation isn't commoditized like cloud storage. A single labeled image might cost $0.10 or $5 depending on whether you need basic bounding boxes or expert medical analysis. The same applies across text, audio, and video labeling. Your costs hinge on three interdependent variables: annotation type, annotator expertise level, and quality assurance overhead.
Most vendors quote per-unit pricing (per image, per sentence, per video minute), but hidden costs emerge when you account for project setup, quality control rounds, and revision cycles. A project quoted at $10,000 can balloon to $18,000 after two QA passes catch inconsistencies.
Typical Pricing Ranges by Task Type
Image annotation (bounding boxes, semantic segmentation) typically runs $0.15–$2 per image for standard computer vision tasks. Specialized medical imaging jumps to $3–$10+ per image due to required annotator credentials.
Text labeling (sentiment classification, named entity recognition, intent detection) falls between $0.05–$0.50 per sentence for straightforward tasks. Complex legal document review climbs to $1–$3 per page when domain expertise is mandatory.
Video annotation is the most expensive category. Expect $5–$50 per minute of footage, depending on whether you need frame-by-frame tracking or simple scene classification.
Audio transcription and labeling costs $1–$5 per minute for standard speech, with specialized dialects or noisy environments pushing higher.
These ranges assume moderate quality standards. Demanding 99%+ accuracy from expert annotators multiplies costs by 1.5–3x.
What Drives Price Variation
- Complexity: Simple binary classification is cheap; multi-label hierarchical categorization isn't
- Expertise required: Crowdsourced workers cost less than PhDs in chemistry reviewing spectroscopy data
- Volume: Labeling 1,000 images costs more per unit than 100,000 due to fixed overhead
- Turnaround time: Rush projects (under 2 weeks) add 20–40% premiums
- Quality standards: Consensus labeling (3+ annotators per item) doubles costs versus single-pass labeling
- Geographic sourcing: US-based annotators cost 3–5x more than Southeast Asian providers; both have quality tradeoffs
Onshore vs. Offshore: The Cost-Quality Tradeoff
US and European annotators typically command $15–$35/hour, translating to higher per-unit costs but fewer revision cycles if your brief is clear. Offshore providers in India, the Philippines, or Vietnam cost $3–$8/hour, making them attractive for large volumes—but expect 1–2 QA passes built into your timeline and budget.
The sweet spot for many teams is hybrid sourcing: offshore annotators handle baseline labeling, onshore experts perform QA and handle edge cases. This cuts costs 30–40% versus full onshore while maintaining acceptable accuracy.
Questions to Ask Vendors Before Committing
Don't just ask "how much per image?" Dig into:
- What's included in their base price—annotation only, or does it include project management and QA?
- How many revision rounds are built in, and what's the cost of additional iterations?
- Do they provide inter-annotator agreement (IAA) metrics so you can assess label quality?
- What's their timeline for 10,000 images at your complexity level?
- Can they handle your specific file formats and integrate with your pipeline?
- What happens if you need 30% more labels mid-project—do rates drop, or do you pay the same?
If you're comparing multiple providers, request quotes on the exact same 100-item sample, then evaluate both cost and quality of their work.
Budgeting Beyond Unit Costs
Allocate your annotation budget as roughly:
- 60–70% for actual annotation labor
- 15–20% for quality assurance and consensus reviews
- 10–15% for project management, revisions, and contingency
If a vendor quotes only the labor portion and ignores QA complexity, they're either underselling or setting you up for poor outcomes.
Frequently Asked Questions
Q: Is crowdsourced annotation cheaper than managed teams, and is it worth the savings? Crowdsourced platforms (Amazon Mechanical Turk, Prolific) cut per-unit costs by 40–60%, but require robust quality controls, more QA iterations, and your own project management overhead. Best for high-volume, lower-stakes tasks; avoid for specialized domains.
Q: How do I know if a quoted price is actually competitive? Request 3–5 vendor quotes on an identical small test set (100–500 items), then evaluate the quality of their work alongside cost. Price alone means nothing; a vendor charging 50% more but delivering usable labels in one QA pass beats a cheaper option requiring three rounds.
Q: Should we hire in-house annotators or outsource? In-house makes sense only for ongoing projects needing 5,000+ labels monthly with proprietary data security requirements. For one-off or episodic labeling, outsourcing avoids fixed salary costs and ramp-up time.
Compare vetted data annotation providers side-by-side on Mercoly to find the right balance of cost and quality for your project.