Data Annotation Pricing In 2027: Per Label, Per Hour And Managed Rates

Meta Description: Data annotation pricing in 2027 explained: per label, per hour and managed rates, cost by data type, platform fees, hidden costs and how to read a quote.

Data annotation pricing in 2027 starts at a few cents per label for simple image work and climbs to hundreds of dollars per hour for physicians, lawyers and senior engineers producing expert training data. Published starting rates sit around $0.02 to $0.10 per bounding box and $5 to $10 per annotator hour for general labeling teams, while managed programs and expert RLHF work are almost always quoted privately. Every vendor prices against its own workforce, quality process and margin, so treat the figures below as published reference points rather than a market index.

That spread is wide enough to be useless on its own. What buyers actually need to know is which pricing model fits the project, which variables move the number, and which costs sit outside the per-label rate. This guide covers all three, using prices published on vendor pricing pages and figures from reputable reporting as of October 2026. If you are still choosing a provider, start with our list of the top data annotation and labeling companies.

Data Annotation Pricing At A Glance

Pricing Model	How You Pay	Published Reference Points	Best Fit
Per label, object or task	A fixed price per box, polygon, entity or item	Bounding boxes from $0.02 (Label Your Data) to $0.10 (Roboflow labeling services)	Well-defined, high-volume tasks with stable guidelines
Per annotator hour	Billed for workforce time	About $6 per hour (Label Your Data), $6.12 vs $28.80 per hour for offshore vs US-based teams (iMerit on AWS Marketplace)	Changing guidelines, mixed tasks, ongoing QA
Managed program or FTE	Monthly or project fee for a dedicated team, PM and QA	Rarely published; CVAT lists a $5,000 minimum project budget	Long-running programs, regulated data, multi-modal work
Expert and RLHF data	Hourly rate for credentialed contributors, plus platform margin	Mercor lists a $93 average contracted rate; Bloomberg reported $110 in October 2026	LLM post-training, evaluation, medical and legal data
Software platform only	Seat, usage or annual subscription; you supply labelers	CVAT from $23 per user per month (yearly); Label Studio Starter Cloud $99 per month	Teams with their own annotators

The Three Main Pricing Models

Per Label, Per Object Or Per Task

Per-unit pricing is the most common model for computer vision and simple text work. You pay for each completed annotation: one bounding box, one polygon, one named entity or one classified image. Label Your Data publishes starting prices of $0.015 per keypoint, $0.02 per bounding box object and $0.02 per entity for NLP tasks. Roboflow lists its labeling services, an Enterprise add-on, starting at $0.05 per classification or keypoint, $0.10 per bounding box and $0.20 per polygon. AnnotationBox lists bounding boxes at $0.04 per object and polygons at $0.06, though its FAQ gives a $0.05 to $0.07 range for 2D boxes.

The appeal is predictability: multiply object count by rate and you have a number. The risk is that the unit rate hides everything around it. A HumanSignal learning center article from June 2026 points out that tooling, QA and rework are often excluded from quoted per-label rates.

Per Annotator Hour

Hourly billing charges for workforce time regardless of how many labels get produced. Label Your Data says it typically charges $6 per annotator hour for flexible workflows or ongoing QA. AnnotationBox lists $5 to $7 per annotator hour. On AWS Marketplace, iMerit has listed its standard data labeling service at $6.12 per hour and its USA-based workforce at $28.80 per hour, both billed in 10-second increments.

Labelbox shows two labeling rates on its pricing page, $8 per hour and “starting at $10 per hour,” without stating exactly which tier each applies to, so confirm before you budget. Hourly pricing works well when guidelines are still moving or tasks are mixed, but it moves productivity risk onto the buyer. If annotators are slow, you pay for it.

Managed Program Or Dedicated FTE Team

For long-running work, many vendors assign a dedicated team with a project manager, QA leads and a fixed monthly or project fee. Rates for this model are rarely published. CVAT is one of the few to state terms publicly: its annotation services page lists a $5,000 minimum project budget, pricing per object, per image or video, or custom, a free proof of concept, discounts of up to 30% for large datasets and a typical turnaround of about one month for contracted projects.

Managed programs cost more per unit of output on paper, but they bundle the guideline work, QA and coordination that per-label quotes often leave out. CVAT’s own July 2024 cost article estimated that outsourcing costs more than 1.5 times a potential in-house team, which is a useful reminder that you are paying for management as well as labor.

What Drives Data Annotation Cost

Data Type And Annotation Complexity

A whole-image classification tag takes seconds. A pixel-accurate segmentation mask can take minutes. Price follows effort, which is why published rates step up from keypoints and boxes to polygons and masks. CVAT describes semantic segmentation as one of the most expensive annotation types.

Object Density

Per-object pricing scales with how crowded your images are. A street scene with 40 pedestrians and vehicles costs many times more than a product photo with one item, even at the same per-box rate.

QA Layers And Accuracy Targets

Consensus labeling, where two or more annotators label the same item, multiplies labor directly. Review passes, gold-standard honeypot tasks and audit sampling add more. CVAT says its services use ground truth, honeypots and consensus checks, plus a second annotator where needed. Ask how many passes are included at the quoted rate.

Domain Expertise

General annotators and credentialed experts are different labor markets. The HumanSignal article cites an estimate from an unnamed AI lab that expert annotators cost 20 to 40 times the hourly rate of generic crowd workers. Medical, legal and code data sit at the expensive end.

Turnaround Time

Rush timelines mean overtime, more parallel annotators and more QA overhead. Vendors price urgency either as a surcharge or by quoting a higher rate overall.

Security And Compliance Requirements

Data that cannot leave your cloud, needs HIPAA handling or requires on-site or clean-room workers narrows the vendor pool and raises the rate. AWS’s Ground Truth documentation warns users not to share protected health information with vendor workforces unless the vendor’s compliance terms allow it, and that kind of restriction usually means a more expensive workforce.

Workforce Location

Location is one of the biggest single swings. The iMerit Marketplace listings show the same company charging $6.12 per hour for its standard workforce and $28.80 per hour for a USA-based team, more than four times as much. Data residency rules or a requirement for native speakers of a specific language can force that choice.

Cost By Data Type

The ranges below combine published vendor starting prices with a per-label comparison table compiled by BasicAI from publicly available vendor prices. BasicAI labels its table “for reference only” and notes that it does not reflect volume discounts. Where we give a typical range, it runs from the lowest to the highest published figure we found, and actual quotes vary with the drivers above.

Data Type	Published Reference Points	Typical Range And How It Was Derived
Bounding boxes	$0.02 (Label Your Data), $0.04 (AnnotationBox), $0.10 (Roboflow) per box	About $0.02 to $0.10 per box from vendor pages; BasicAI’s compiled table runs $0.03 to $1.00
Polygons	$0.06 (AnnotationBox), $0.20 (Roboflow) per polygon	About $0.05 to $0.26 per polygon, from vendor pages and BasicAI’s table
Semantic segmentation masks	$0.05 to $3.00 per mask in BasicAI’s compiled table	Wide range; dense or medical masks sit at the top
LiDAR and 3D cuboids	$0.05 to $3.00 per cuboid in BasicAI’s compiled table	Wide range; frame count, point density and tracking drive it
Video	$0.50 to $10.00 per minute in BasicAI’s guide	Per-frame or per-track pricing varies by vendor
Text NER and classification	$0.02 per entity (Label Your Data)	Low cents per entity for general text; more for legal or clinical text
Audio transcription	$1.99 per minute for human transcription (Rev)	BasicAI’s guide gives $0.10 to $10.00 per audio minute
Medical imaging	$2.00 to $8.00 per image for medical or scientific segmentation in BasicAI’s guide	Specialist time dominates; see expert rates below
RLHF and expert preference data	$93 average contracted rate (Mercor), $110 average (Bloomberg, October 2026)	Hourly, from about $50 for generalists to $400 for specialist attorneys

Bounding Boxes And Keypoints

These are the cheapest image annotations and the most widely published. Label Your Data starts at $0.02 per box and $0.015 per keypoint. Roboflow’s labeling services start at $0.10 per box and $0.05 per keypoint. The roughly fivefold gap between those two starting prices shows how much vendor model and quality process matter even on the simplest task.

Polygons And Segmentation

Polygons trace an object’s outline, and masks label every pixel. In BasicAI’s compiled table, polygon prices run from $0.045 to $0.257 and segmentation masks from $0.05 to $3.00 depending on the vendor. Budget well above box rates if your model needs exact boundaries.

LiDAR, 3D Cuboids And Sensor Fusion

Autonomous vehicle and robotics data is labeled in 3D point clouds, often across sequences of frames with objects tracked between them. BasicAI’s table lists cuboid prices from $0.05 to $3.00 per label across vendors. Most specialist 3D vendors, including BasicAI and Label Your Data for cuboids, do not publish a rate and price after a pilot.

Video

Video can be priced per frame, per tracked object or per minute. BasicAI’s guide puts video at $0.50 to $10.00 per minute. Interpolation tools that fill frames between keyframes cut the cost considerably, so ask whether the quote assumes manual frame-by-frame work.

Text And NER

Named entity recognition, sentiment and classification are usually priced per entity or per document. Label Your Data starts at $0.02 per entity. Clinical, legal and financial text costs more because annotators need subject knowledge.

Audio Transcription

Rev lists human transcription at $1.99 per minute, with foreign-language subtitles from $6.49 to $15.99 per minute depending on the language. Speaker diarization, timestamps and noisy recordings add cost on top of plain transcription.

Medical Imaging

Medical annotation splits into two costs: the per-image work and the clinician time behind it. BasicAI’s guide puts medical or scientific segmentation at $2.00 to $8.00 per image. Built In reported in November 2025 that Mercor roles for primary care physicians paid $130 to $170 per hour, and Mercor’s homepage currently lists an oncologist role at $350 to $450 per hour.

RLHF And Expert Preference Data

Post-training data for large language models is priced by the hour of expert time, not by label. Mercor lists a $93 average contracted rate on its homepage, and Bloomberg Law reported on October 16, 2026 that average pay at Mercor had reached $110 per hour, with civil rights attorneys earning $400 per hour and software engineers up to $120 on AI training work. DataAnnotation advertises up to $50 per hour for generalists, $125 or more for finance, legal and medical experts and $150 or more for coding.

These are contributor pay rates, not what buyers pay. TechCrunch reported in September 2025 that Mercor charges customers an hourly finder’s fee and matching rate on top of contractor pay. On self-serve platforms, Prolific lists a minimum participant pay of $8 per hour, recommends at least $12, and adds a platform fee that is usually 42.8% for corporate customers. Compare vendors in our RLHF and LLM training data companies list.

Software Platform Pricing Models

If you have your own annotators, you buy software instead of labor. Platforms charge in four main ways: per seat, per usage hour, by credits or by annual contract. Our AI data annotation platforms list compares features in more depth.

Platform	Published Price (October 2026)	Model
CVAT Online	Solo $33 per month or $23 per month yearly; Team $33 per user per month or $23 yearly; free Community edition self-hosted	Per seat
Label Studio (HumanSignal)	Starter Cloud $99 per month, $49 per month per additional user, up to 12 users; open-source Community edition free	Per workspace plus seats
Roboflow	Free tier with 10 credits per month; Core $39 per month; prepaid credits from $4, on-demand $6 per credit	Subscription plus credits
Supervisely	Pro from €199 per month; labeling seats €35 per user; labeling hours €0.65 per hour	Subscription plus seats or hours
Segments.ai	Core starts at 3,600 labeling hours per year, quoted as equivalent to $2.67 per hour	Usage hours, custom quote
BasicAI	Private-cloud deployment from $6,600 per year	Annual license
Labelbox, Encord, Kili, SuperAnnotate, V7	Enterprise pricing through sales	Custom

HumanSignal, Supervisely and Segments.ai publish enough to budget against. For the custom-priced platforms, ask whether the quote is per seat, per asset or per usage unit, because the model changes how costs grow with your dataset.

A Note On Amazon SageMaker Ground Truth

Buyers who budgeted against AWS should check their plans. Amazon SageMaker Ground Truth documentation now states that the service is no longer open to new customers and will receive no new features. Amazon Mechanical Turk permanently closed on September 30, 2026, and its help page says the closure also applies to the Mechanical Turk workforce in Ground Truth and Amazon Augmented AI. Existing customers can still use private and vendor workforces, but new projects need another route.

The Costs Nobody Puts In The Quote

Guideline Creation

Someone has to write the labeling instructions, edge-case rules and examples. That usually falls on your ML engineers or data leads, and it is rarely in the vendor’s price. Poor guidelines are the main cause of rework later.

Pilots And Calibration

Most vendors offer a free pilot. Label Your Data, CVAT and BasicAI all advertise one. The pilot is free in dollars but not in your team’s time spent preparing samples, reviewing output and revising guidelines.

Rework

Per-label quotes usually price a single pass. If a guideline changes halfway through, relabeling may be billed again. Ask who pays for rework caused by vendor error versus rework caused by changed requirements.

Project Management And Internal Review

You still need someone to own the vendor relationship, spot-check deliveries and resolve questions. The HumanSignal article estimates that labor makes up 60% to 80% of total labeling cost, tooling 10% to 15% and QA 15% to 20%, which shows how much sits outside raw annotation.

Tooling Seats And Storage

If the vendor works in your platform, you may pay for their seats. Supervisely, for example, charges €35 per labeling user or €0.65 per labeling hour on Pro, plus storage add-ons. Model-assisted labeling can also consume compute credits, as on Roboflow.

A Worked Example: 50,000 Images Of Street Scenes

Assume a project with 50,000 images averaging eight objects each, or 400,000 bounding boxes. The figures below use published starting rates; the hourly scenario uses an assumed throughput of 100 boxes per annotator hour purely for illustration, and your pilot should measure the real number.

Scenario	Calculation	Labor Cost
Per label at $0.02 (Label Your Data starting rate)	400,000 × $0.02	$8,000
Per label at $0.04 (AnnotationBox listed rate)	400,000 × $0.04	$16,000
Per label at $0.10 (Roboflow labeling services starting rate)	400,000 × $0.10	$40,000
Per hour at $6 (Label Your Data typical rate)	4,000 hours × $6	$24,000
Per hour at $28.80 (iMerit USA-based listing)	4,000 hours × $28.80	$115,200
Consensus labeling, two annotators per image at $0.04	800,000 × $0.04	$32,000

Then add the costs outside the quote: internal hours for guidelines and the pilot, review of a sample of each delivery, a rework allowance and any tool seats. A self-managed alternative on CVAT Team, with 10 of your own annotators on yearly billing for three months, would cost about $690 in seats, but your annotators’ wages become the main cost.

For a published reference at larger scale, CVAT’s July 2024 cost article worked through 100,000 images containing about 2.3 million objects and estimated approximately $225,400 at its highest volume discount tier, or just under $0.10 per object. Starting rates are a floor, not an average.

How To Reduce Data Annotation Cost

Use Model-Assisted Prelabeling

Have a model draft labels and let humans correct them. AWS’s Ground Truth documentation says automated labeling reduces cost and time compared with humans alone, and recommends at least 5,000 objects for it to work well. Most modern platforms offer similar features with foundation models.

Match Expertise To The Task

Do not pay expert rates for work generalists can do. Route easy items to general annotators and escalate hard cases to specialists.

Sample QA Instead Of Reviewing Everything

Full double-labeling doubles cost. Statistical sampling plus honeypot tasks catch most quality problems for a fraction of the price.

Lock Guidelines Before Scaling

Run the pilot until edge-case questions stop, then scale. Changing rules at 300,000 labels is far more expensive than at 3,000.

Negotiate Volume Tiers

Published discounts exist. CVAT lists up to 30% off for large datasets. Commit to volume only after the pilot proves quality.

Label Less, Label Smarter

Active learning and deduplication cut the number of items that need human attention. Many datasets contain near-duplicate frames that add cost without adding information.

How To Read A Data Annotation Quote

Line up quotes on the same basis before comparing them. Check these points:

• Unit definition. Is the price per object, per image, per frame, per minute or per hour? A per-image price for crowded scenes can beat a per-object price, or the reverse.
• QA included. How many review passes, what accuracy target and what happens if a delivery misses it?
• Rework terms. Who pays to fix vendor errors and changed requirements?
• Minimums and setup. CVAT, for example, lists a $5,000 minimum. Some vendors charge setup or guideline fees.
• Tooling. Whose platform is used, and are seats or storage billed separately?
• Workforce. Where are annotators located, are they employees or contractors, and what security controls apply?
• Turnaround. Is the timeline a commitment, and is there a rush surcharge?
• Payment terms. Pay on completion, pay upfront for credits, or monthly retainers.

For broader budgeting beyond data, our AI development cost guide covers model build, infrastructure and maintenance costs.

Frequently Asked Questions

How Much Does Data Annotation Cost Per Image?

It depends on how many objects each image contains and the label type. BasicAI’s compiled guide gives $0.01 to $5.00 per image overall, with basic classification at $0.03 to $0.10 and medical or scientific segmentation at $2.00 to $8.00. For object detection, multiply the object count by a per-box rate.

Is Per-Label Or Per-Hour Data Labeling Cost Lower?

Neither is cheaper in every case. Per-label pricing rewards stable, high-volume tasks because the vendor carries productivity risk. Per-hour pricing suits changing tasks but leaves you paying for slow work. Run a pilot and convert both quotes to cost per accepted label.

Why Is RLHF Data So Much More Expensive Than Image Labeling?

RLHF requires people who can judge model output in a domain: doctors, lawyers, engineers or skilled writers. Mercor lists a $93 average contracted rate for its experts, and Bloomberg reported specialist attorneys earning $400 per hour, compared with a few dollars per hour for general image annotation.

Do Vendors Offer Free Pilots?

Many do. Label Your Data, CVAT and BasicAI all advertise free pilots or proofs of concept. Use the pilot to measure throughput and accuracy, not just to see sample labels.

Is Open-Source Annotation Software Really Free?

The license is. CVAT Community and Label Studio Community Edition can be self-hosted without fees, but you pay for hosting, maintenance, security and the annotators themselves.

Are Published Rates What I Will Actually Pay?

Usually not exactly. Published figures are mostly starting prices for simple tasks at volume. Expect quotes to move with complexity, QA, security and turnaround, and confirm current pricing with each vendor, since rates change often.

Conclusion

Data annotation pricing in 2027 runs from about two cents per bounding box to several hundred dollars per expert hour, and the right model depends on your data type, how settled your guidelines are and how much quality control you need. Published per-label rates are a starting point, not a budget.

To build a realistic number, choose a pricing model, run a paid or free pilot to measure real throughput and accuracy, then add guideline work, QA, rework, project management and tooling. Compare vendors on cost per accepted label rather than headline rate, and check a shortlist of established providers before you request quotes.