How Much Does Data Labeling Cost? Pricing Models Explained

By Stefan, founderBasics5 min read

Data labeling cost comes down to one formula: the time the work takes, multiplied by the rate for the people doing it, plus the cost of checking it. The time depends on how hard each item is, how much detail is needed and how much checking the quality level requires. That is why two quotes for "10.000 items" can differ by many times and both be honest.

This guide explains the three pricing models, what drives the price, how to estimate a project yourself and how to compare quotes without being misled by a low headline number.

The Three Pricing Models

Model You pay for Works best for Watch out for
Per hour Time annotators and reviewers spend Tasks where item length varies a lot: AI answer evaluation, expert review, long documents Ask for weekly reports of hours and output, so costs stay predictable
Per item (per task) Each labeled item Simple, repeatable items of similar size: tags, short texts, bounding boxes A low per-item rate can push annotators to rush hard items
Per project (fixed price) An agreed scope, delivered Well-defined, one-off datasets Changes in scope mean a new quote; the vendor prices in risk

Per hour is the most transparent: you see exactly what you pay for and annotators are not pushed to rush. It needs good reporting to stay predictable.

Per item gives a clear total upfront. It works well once the task is stable and the speed is known, usually after a pilot.

Fixed price suits clear, one-off jobs. Expect a margin for the vendor's risk if the work turns out harder than planned.

A common path is to pilot on an hourly basis, measure the real speed, then switch to a per-item price for the full volume.

What Drives the Price

Cost driver Why it matters Cheaper end More expensive end
Task difficulty Harder judgments take longer per item One tag per short text Comparing two long AI answers and explaining why
Detail per item More marks per item means more time Image-level tag Pixel-level segmentation
Expertise Professionals cost more per hour General reading skills Lawyers, doctors, engineers
Language Rare languages narrow the pool of annotators Widely spoken languages Less common languages, native speakers required
Quality level Overlap and review add hours Spot checks only Double labeling plus full review
Guideline clarity Unclear rules cause questions and rework Clear, tested guidelines Rules that change every week
Volume and deadline Rush jobs need more people at once Steady volume, normal deadline Large volume, very short deadline
Security needs Restrictions limit who and how Anonymised data Restricted countries, special access setup

Of these, guideline clarity is the one you control most directly. Clear rules make every other cost lower.

A Worked Example: Estimating a Project Yourself

The numbers below are illustrative only, not a quote. They show how to build an estimate; your own speeds and rates come from a pilot.

The project: classify 10.000 short customer emails into 8 topics. A pilot showed annotators handle about 120 emails per hour and reviewers check about 240 per hour.

Work Items Speed Hours
Labeling 10.000 120 per hour 83.3
Overlap: 10% labeled a second time to measure agreement 1.000 120 per hour 8.3
Review: 10% checked by a reviewer 1.000 240 per hour 4.2
Total 95.8

At an example all-in rate of $20 per hour, the project costs about $1.917, or about $0,19 per email.

How the estimate changes:

  • If the emails were long and needed highlighted reasons, speed might halve and the cost would roughly double.
  • If the topic labels were unclear and 15% of the work had to be redone, you would pay for that rework too.
  • Dropping the overlap and review would cut about 13% of the hours, but you would no longer know your quality.

The lesson: the pilot's measured speed is the most important number in any estimate. Without it, every quote is a guess.

How to Compare Quotes Fairly

A low headline price can hide high real costs. When comparing vendors, ask each one the same questions:

  • What exactly is included? Labeling only, or also overlap, review, project management and quality reports?
  • Who pays for rework? If a batch fails the agreed accuracy, is fixing it free?
  • What accuracy do you commit to and how is it measured?
  • What are the annotators paid? Very low pay usually means high turnover, constant retraining and uneven quality.
  • Are there setup fees, minimum volumes or tool fees?
  • Is any of the work done by AI? Machine pre-labels checked quickly by people can look cheap, but errors pass through.

Then compare on cost per correctly labeled item: the total price divided by the number of items that pass your own checks. A cheaper quote with more errors often ends up costing more, once you add your own time spent finding and fixing mistakes.

Hidden costs on your side count too: time to write and update guidelines, answer questions, check samples and manage the vendor. Good vendors reduce that time; poor ones add to it.

How to Lower the Cost Without Losing Quality

  1. Write clear guidelines and test them. The cheapest hour is the one not spent on rework.
  2. Use the simplest annotation that serves the goal. A tag instead of a box, a box instead of an outline, where the model allows.
  3. Merge labels the model doesn't need to tell apart.
  4. Remove items that don't need labeling, such as duplicates and empty records, before you start.
  5. Label the hard cases first. Let the model show you where it is weak and focus there. See Human-in-the-Loop AI.
  6. Match the quality level to the use. A test set deserves full review; a large first training set may need less.
  7. Run a pilot. It turns guesses into measured speed and cost.

See the All-In Price Before You Commit

At AnnoTrue, the price is the annotator's pay plus a service fee, nothing hidden. Our pricing calculator shows the all-in price for your project type and founding clients get preferential pricing locked for 12 months. Start with a free pilot to measure the real speed on your data.