Buying Guide / 6 min read

Data Annotation Pricing: Cost Models and What Drives Cost

VIPRO Journal30 Oct 2026 · Buying Guide

Buyer's guide

Short answer: data annotation is usually priced per object or label, per item (image, minute of video, document), per hour of annotator time, or as a fixed project fee. The final cost depends far more on task detail, edge-case rules and quality requirements than on the pricing model itself. A short paid or free pilot on your own data is the most reliable way to get a realistic number.

The four common pricing models

Model	How it works	Works best when	Risk to watch
Per object / per label	You pay for each box, polygon or tag	Object counts are predictable	Dense scenes get expensive fast
Per item	Fixed price per image, video minute or document	Items are similar in difficulty	Hard items are underpriced, easy ones overpriced
Per hour	You pay for annotator and reviewer time	Scope is still changing or tasks are complex	Needs clear productivity reporting
Fixed project	One price for an agreed scope	Scope, rules and volumes are stable	Change requests can be costly

No model is "cheapest". The right one matches how predictable your data is.

What actually drives the cost

Annotation type. A class tag is quicker than a box, a box is quicker than a polygon, and a polygon is quicker than a pixel mask. 3D cuboids and video tracking add further effort.
Objects per item. A highway frame with 40 vehicles costs more than an empty road, even at the same "per image" price.
Edge-case rules. Occlusion, truncation, blur, night scenes and rare classes slow work down and need extra review.
Domain expertise. Medical, legal or industrial data may need trained or expert reviewers.
Quality target and review depth. A second full review pass, gold-set checks or agreement measurement improve quality but add time.
Tooling and formats. Working in your own tool, or converting to a custom export format, can add setup effort.
Security requirements. NDAs, restricted access, on-premise or VPN-only work, and data deletion rules change how a team operates.
Volume and timeline. Larger, steady volumes allow a stable trained team; rushed deadlines need more people at once.

Hidden costs to ask about

Rework: who pays when labels fail your QA?
Rule changes: how are guideline updates handled mid-project?
Project management and reporting: included or extra?
Minimum order or setup fees.
Your own team's time spent answering questions and reviewing samples.

Why the cheapest quote can cost more

Low-quality labels are paid for twice: once to create them and again to find and fix them, often after a model has already trained on them. Inconsistent labels also make it hard to tell whether a model problem comes from the model or from the data.

Questions to ask before you sign

Which pricing model are you proposing, and why does it fit my data?
What is included: review passes, QA reports, project management, format conversion?
How do you measure quality, and what happens when a batch fails?
Can we run a pilot on my own data before committing to volume?
How are guideline changes priced?
What security controls apply to my data, and how is it deleted at the end?

How a pilot helps you price accurately

A pilot uses a few hundred real items and your actual rules. It shows real throughput, the edge cases that slow work down, and the quality level you can expect, so the full quote is based on evidence rather than guesses. VIPRO starts most projects with a calibration pilot for this reason.

Have a dataset in mind?

Start with a Free Calibration Pilot