Data labeling is often the most expensive part of building an ML model, yet most teams underestimate the cost. SageMaker Ground Truth provides three workforce options — Amazon Mechanical Turk, private teams, and third-party vendors — each with different quality and cost trade-offs. At scale, labeling 100,000 images with bounding boxes costs $3,600 or more through Mechanical Turk alone.
The real cost saver is automated labeling with active learning. Ground Truth trains an intermediate model on your human-labeled data and uses it to auto-label high-confidence items, reducing the number of items sent to human annotators by up to 70%.
TL;DR: Ground Truth label pricing varies by type: image classification at $0.012/image, bounding boxes at $0.036/object, semantic segmentation at $0.07/image (Mechanical Turk rates). Active learning reduces labels needing human review by up to 70%, cutting costs proportionally. For 100K image classification labels, expect $1,200 without active learning or $360-$500 with it.
Labeling Pricing by Type
Amazon Mechanical Turk Pricing
Mechanical Turk provides the most affordable per-label rates. Each data object is labeled by 3-5 workers by default for quality consensus.
Image labeling:
| Label Type | Price per Object | 10K Objects | 100K Objects |
|---|---|---|---|
| Image Classification | $0.012/image | $120 | $1,200 |
| Multi-label Classification | $0.012/image | $120 | $1,200 |
| Bounding Box | $0.036/object | $360 | $3,600 |
| Semantic Segmentation | $0.07/image | $700 | $7,000 |
| Instance Segmentation | $0.084/image | $840 | $8,400 |
| Image-level Labeling (Polyline) | $0.036/object | $360 | $3,600 |
Text labeling:
| Label Type | Price per Unit | 10K Units | 100K Units |
|---|---|---|---|
| Text Classification | $0.012/unit | $120 | $1,200 |
| Named Entity Recognition | $0.012/unit | $120 | $1,200 |
| Multi-label Text Classification | $0.012/unit | $120 | $1,200 |
Video and 3D Point Cloud labeling:
| Label Type | Price per Unit |
|---|---|
| Video Object Detection | $0.036/frame |
| Video Object Tracking | $0.048/frame |
| 3D Point Cloud Object Detection | $0.30/frame |
| 3D Point Cloud Segmentation | $0.468/frame |
Private Workforce
The Ground Truth platform is free when using a private workforce. You pay your workers directly through your own payroll or contractor agreements.
| Component | Cost |
|---|---|
| Ground Truth platform | Free |
| Labeling UI and tools | Free |
| Worker management | Free |
| Worker compensation | You pay directly |
Private workforces are ideal for sensitive data that cannot be shared externally, domain-specific tasks requiring specialized knowledge, or when you have in-house annotators.
Third-Party Vendors
AWS Marketplace vendors provide managed labeling teams with quality guarantees. Pricing varies by vendor and task complexity:
| Vendor Type | Typical Price Range | Best For |
|---|---|---|
| Standard annotation | 2-5x Mechanical Turk rates | Higher quality, managed QA |
| Specialized (medical, legal) | 5-20x Mechanical Turk rates | Domain expertise required |
| Enterprise managed | Custom pricing | Large-scale, ongoing projects |
Active Learning (Automated Labeling)
Active learning is Ground Truth's most powerful cost optimization feature. It works in two phases:
- Human labeling phase: Workers label an initial batch of data (typically 1,000-5,000 items)
- Automated labeling phase: Ground Truth trains a model and auto-labels items where it has high confidence
| Dataset Size | Without Active Learning | With Active Learning (70% auto) | Savings |
|---|---|---|---|
| 10,000 images (classification) | $120 | $36 + $4 compute = $40 | 67% |
| 50,000 images (classification) | $600 | $180 + $15 compute = $195 | 68% |
| 100,000 images (classification) | $1,200 | $360 + $30 compute = $390 | 68% |
| 100,000 images (bounding box) | $3,600 | $1,080 + $30 compute = $1,110 | 69% |
The compute cost for active learning is the ML training and inference needed to build and run the auto-labeling model. This is typically $2-$5 per 10,000 data objects — negligible compared to human labeling savings.
Requirements for active learning:
- Minimum 1,000 labeled objects to start
- Works best with classification and bounding box tasks
- Confidence threshold is configurable (default 95%)
- Not available for all label types (e.g., not for 3D point cloud)
Ground Truth Plus
Ground Truth Plus is a fully managed labeling service where AWS provides the expert workforce and project management.
| Feature | Ground Truth | Ground Truth Plus |
|---|---|---|
| Workforce management | You manage | AWS manages |
| Quality control | You configure | AWS handles |
| Project management | Self-service | Dedicated PM |
| Pricing model | Per-label | Custom contract |
| Setup effort | DIY | Turnkey |
| Domain expertise | Depends on workforce | Specialized teams available |
Ground Truth Plus typically costs 3-5x more per label than self-managed Mechanical Turk, but eliminates the operational overhead of managing labeling projects. It is best for enterprises that need high-quality labels at scale without building an internal annotation operations team.
Real-World Cost Scenarios
Computer Vision Startup (Object Detection)
| Component | Details | Cost |
|---|---|---|
| Initial labeling | 20,000 images, bounding boxes, MTurk | $720 |
| Active learning enabled | 80,000 images auto-labeled | $120 compute |
| Quality review | 5,000 samples re-labeled | $180 |
| S3 storage | 100K images, 50GB | $1.15 |
| Total for 100K labeled images | $1,021 |
NLP Team (Text Classification)
| Component | Details | Cost |
|---|---|---|
| Text classification | 50,000 documents, MTurk | $600 |
| Active learning | 150,000 documents auto-labeled | $50 compute |
| NER labeling | 20,000 documents, private workforce | $0 (platform) + labor |
| Total (excl. private labor) | $650 |
Autonomous Driving (3D Point Cloud)
| Component | Details | Cost |
|---|---|---|
| 3D object detection | 10,000 frames, vendor | $9,000 |
| Video object tracking | 50,000 frames, vendor | $7,200 |
| Quality review and re-labeling | 10% re-work | $1,620 |
| Total | $17,820 |
Cost Optimization Tips
-
Enable active learning for every eligible job. It reduces human labeling volume by up to 70%, cutting costs proportionally. The compute overhead is negligible — $2-5 per 10,000 items.
-
Start with a small labeled sample to validate your task design. Label 500-1,000 items first, review quality, refine instructions, then scale. Poorly designed tasks waste money on low-quality labels that need re-work.
-
Use consensus wisely. The default is 3 annotators per item. For simple tasks (binary classification), 3 is sufficient. For complex tasks (segmentation), 5 may improve quality. For high-confidence active learning labels, consensus is handled automatically.
-
Pre-filter your data before labeling. Remove duplicates, irrelevant images, and corrupted files before submitting labeling jobs. Every item in your dataset incurs a per-label cost.
-
Use private workforces for sensitive or specialized data. The Ground Truth platform is free for private workforces. If you have domain experts (radiologists for medical imaging, lawyers for legal documents), leverage them without per-label platform fees.
-
Batch your labeling jobs. Larger batches are more cost-efficient with active learning because the auto-labeling model improves with more training data. A single job of 100,000 items is cheaper than ten jobs of 10,000 items.
Related Guides
- AWS SageMaker Pricing: Training, Inference, Studio
- AWS SageMaker Training Guide
- AWS S3 Pricing Guide
FAQ
How much does it cost to label 100,000 images?
The cost depends on label type and workforce. Image classification on Mechanical Turk: $1,200. Bounding boxes: $3,600. Semantic segmentation: $7,000. With active learning enabled (70% auto-labeled), these costs drop to approximately $390, $1,110, and $2,130 respectively. Private workforce and vendor pricing varies.
What is the difference between Ground Truth and Ground Truth Plus?
Ground Truth is a self-service platform — you configure labeling jobs, manage workers, and handle quality control. Ground Truth Plus is a fully managed service where AWS provides expert annotators and project management. Plus costs 3-5x more per label but eliminates operational overhead. Choose Plus for large-scale projects where managing annotators is not your core competency.
How does active learning reduce labeling costs?
Active learning trains an ML model on your initial human-labeled data (typically 1,000-5,000 items). It then uses this model to auto-label items where its confidence exceeds a threshold (default 95%). Only low-confidence items are sent to human annotators. In practice, this auto-labels 40-70% of your dataset, reducing human labeling costs proportionally.
Lower Your SageMaker Ground Truth Costs with Wring
Wring helps you access AWS credits and volume discounts to lower your SageMaker Ground Truth costs. Through group buying power, Wring negotiates better rates so you pay less per labeling task.
