Data Annotation Companies and 12 Labeling Vendors Checked on Published Pricing

We opened twelve data annotation companies in September 2026 looking for a price. Two of them published one. Four had no pricing page that resolved at all, and one category guide states the position out loud: most annotation companies do not publish rates, and sales pages say contact us for a quote. Where numbers do exist they are wilder than the silence suggests, with the same nominal task priced sixty times apart between two well known vendors. Here is what each company actually discloses, the arithmetic that makes incomparable quotes comparable, and the one question the category answers two opposite ways in public.

Do data annotation companies publish pricing?

Mostly no, and the split is not random. The two providers that published a rate card, Label Your Data and BasicAI, are both platform led with a self serve tier, so a public number is part of the product. The providers selling managed annotation teams, HiTech BPO, Anolytics and ARDEM among them, publish nothing and route every pricing question to a proposal form. That is defensible, because a managed team is priced on scope rather than on units, but it means a buyer comparing a platform quote against an agency quote is comparing two different commercial models before they compare a single dollar.

The more useful signal is what happens when a big name is asked directly. Scale AI publishes no unit rate at all, offering the first 1,000 labeling units and 10,000 curated images at no cost and then pay as you go by card. Labelbox, Appen, iMerit and Surge AI all had pricing URLs that returned errors to automated reading, so we have left them as unread rather than guessed at. Four unreadable pricing pages out of twelve is itself a finding about how this category prefers to sell.

What twelve data annotation companies actually disclose

Ranking these firms one to twelve would be invented precision, because no outsider can measure their output quality. What is checkable is what each prints on its own site, read directly in September 2026.

One column deserves a second look. An American address is not an American delivery location. HiTech BPO lists New York, California and London addresses with its headquarters in Ahmedabad, and Anolytics lists a Levittown, New York headquarters with all three of its delivery centers in Noida. Neither is hiding anything, both print it on the site. But it is the same pattern we found when we audited where AI development companies actually build, and it is worth getting in writing rather than inferring from a contact page.

How much do data annotation services cost per label?

Published per unit rates run from $0.015 for a keypoint to $5.00 and up for a medical label. For simple image work the real band is narrow and cheap: classification $0.02 to $0.15, bounding boxes $0.02 to $0.90, with the two published vendor rates sitting right at the floor at $0.02 and $0.03. Segmentation runs $0.50 to $2.00 and up. Audio is $0.10 to $10.00 a minute and video $0.50 to $10.00 a minute.

The instructive part is the extremes on identical nouns. A polygon is $0.05 at BasicAI and $0.257 at Google. A 3D cuboid is $0.05 at BasicAI and $3.00 at Amazon, a sixty times gap. Nobody is being dishonest. A cuboid on a clean synthetic scene and a cuboid on a sparse night time LiDAR sweep are the same word and completely different work, and the word is what buyers search on. Volume moves it further: discounts of 10 to 30 percent are normal above 100,000 units, and one worked example applies 5 to 30 percent to reach roughly $225,400 for 2,300,000 objects across 100,000 images, just under $0.10 an object.

What hourly wage does a per label quote assume?

This is the calculation that makes annotation quotes comparable, and we have not seen a vendor publish it. Multiply the per unit rate by realistic throughput and you get the hourly labor the quote is priced on. Published benchmarks put an experienced annotator at 200 to 400 bounding boxes an hour on clean scenes, and 50 to 150 an hour on dense scenes or where every item is double checked.

Notice that Label Your Data's two published numbers agree with each other. Their $0.02 box at 300 boxes an hour implies $6, which is exactly the hourly rate they also print. That is a vendor being internally consistent, and it tells you plainly what labor market the price is built on. Now set it against the third party benchmark: a directory reviewing 345 annotation companies across 61 countries reports a median hourly rate of $37. A $37 an hour quote and a $0.02 per box quote are not two bids for the same job. They are two different labor markets wearing the same product name.

Two caveats keep this arithmetic honest. The throughput benchmark is specific to bounding boxes, so applying it to segmentation or medical review would produce nonsense; no equivalent public throughput figure exists for those tasks, which is exactly why their rates vary sixty fold. And implied hourly labor is not the vendor's cost, since tooling, QA, project management and margin all sit on top of it.

Is it cheaper to outsource data annotation or hire an in-house team?

The category answers this two opposite ways, in public, and both answers are internally consistent. One published worked example puts 20,000 images at $15,000 to $18,000 with an in house team against roughly $2,600 outsourced, so outsourcing wins by nearly seven times. A different guide tells buyers to expect the outsourcing price to be more than 1.5 times the cost of a potential in house team, so in house wins. They differ on one buried assumption: whether the in house team stays busy.

That is the entire decision. A salaried annotator costs the same whether they label 200 items or 20,000, so hiring is a bet on utilization, and the first example loses that bet by charging a single 20,000 image project for a full time hire's idle capacity. Framed properly, the crossover is not a volume at all, it is a price per unit. An annotator at the $37 median, working at 150 to 400 units an hour, costs about $0.09 to $0.25 per unit. Below that number, buying per label is cheaper. Above it, a dedicated team is.

Applied to real work, simple image tasks at $0.02 to $0.15 sit under the crossover, usually well under, so buy them per label and do not overthink it. Segmentation, medical imaging and expert preference data at $0.50 to $8.00 sit far above it, so a dedicated annotation team that learns your guidelines is normally cheaper and always more consistent than paying per mask.

How do you choose a data annotation company?

Price the task, not the category. Label 200 representative items yourself first, including the awkward ones, and treat that as your ruler. Send the same 200 item paid pilot to every vendor with your guidelines attached, then score the returns against it. That costs a few hundred dollars and replaces the whole vendor selection problem with a measurement instead of a sales call.

Four things belong in every request. Ask for one all inclusive number covering annotation, QA, project management, tooling, rework and reporting, because vendors who quote only the drawing step are quoting a fraction of the job. Ask how many review passes the rate buys, since a $0.02 box needing a second pass at $0.02 is a $0.04 box. Ask for a measured agreement threshold on your gold set, with failed batches re done at no charge, which is worth more than a 20 percent discount. And ask for the delivery country per person in writing, then check it against the addresses on the vendor's own site.

If you are staffing the work yourself rather than buying it, the same gold set does double duty as a hiring filter: scoring every candidate annotator on identical items before they ever touch production data is far more predictive than a resume, and running that as a scored assessment you can auto grade turns a subjective judgment into a number you can rank on.

Which annotation route fits which buyer

Three routes, and the honest version of who each one suits. Buy per label from a published rate vendor when your dataset is a large pile of simple images labeled once. Nothing beats commodity pricing on a commodity task, and at $0.02 to $0.05 an object you should not be reading about dedicated teams at all.

Use a self serve platform when you want to keep the labeling loop inside your own engineering process, iterate on guidelines weekly, and have someone on staff willing to own project management. You will save on margin and spend it in your own team's hours, which is a good trade only if those hours are cheap.

Hire a dedicated team when the work is above the crossover: regulated data where you need to name the people and the country, subjective or expert judgment such as RLHF and evaluation sets, or a continuous pipeline where the same annotators should still be on the project in a year. In that case annotation has stopped being a purchase and become part of the model team, sitting next to the machine learning engineers who consume the labels and the computer vision development work that turns them into something that ships.

What to put in the annotation contract

Six clauses cover the failure modes we have described. A measurable quality threshold on an agreed gold set. Rework priced at zero below that threshold. The delivery country and a rule on subcontracting. Ownership of both the labels and the annotation guidelines you paid to develop, named explicitly rather than assumed under a general work product clause. An export path in an open format on demand, not only at contract end. And a stated policy on whether your data may be reused to train anything.

That last one matters more each year. Annotation is one of the few AI workflows where a human being sees your raw records one at a time, which makes it a data processing arrangement wearing a services contract. If the records are customer data, medical data or anything you would need to disclose about in a breach, the security terms deserve more attention than the rate.

The short version

Only two of twelve data annotation companies publish a price, so most shortlists are built on sales conversations rather than numbers. Where rates exist they run from $0.015 to $5.00 and the same nominal task can differ sixty fold, which means the words on the quote do not define the work. Convert every per unit quote into implied hourly labor before you compare anything. And decide between buying per label and hiring a team on the crossover, roughly $0.09 to $0.25 a unit, rather than on a vendor's worked example that quietly assumed the answer.