Cost Per Usable Label: How to Compare Data Annotation Quotes

The cheapest annotation quote is often the most expensive one you can sign. In the example below, a $0.04-per-label quote ends up costing $0.214 per usable label once your team's review time and rework are counted, while a $0.06 quote with a quality guarantee costs $0.066. That is a 3.3x difference, in the opposite direction from the sticker price.

The number to compare is cost per usable label: what you pay, in vendor fees and your own team's hours, for a label that survives your QA. This guide shows how to calculate it, includes a calculator you can run on your own quotes, and lists the line items vendors leave out.

Why is price per label the wrong number?

Because the vendor's invoice is only one of three costs, and often the smallest. The other two land on your own team:

• Review. Someone on your side checks the labels. The worse the vendor, the more you must check, because you can't tell which labels are wrong without looking.
• Rework. Rejected labels go back for another pass. Some vendors bill that pass again. Either way, your team reviews it again.

Put together:

The expensive input is usually the loaded hourly cost of whoever does the reviewing. On most AI startups that's an ML engineer. US AI engineers earned $140K–$185K base in 2026, with total compensation commonly above $200K, so an hour of their time costs far more than any per-label fee. Every hour they spend checking boxes is an hour not spent on the model.

What are the three pricing models, and what do they really charge for?

Almost every annotation quote uses one of three models. Each one shifts a different risk onto you.

Public price ranges are wide because task complexity varies by more than 30x. Drawing a box is not segmenting a lane line, and transcribing clean audio is not diarizing a four-person sales call. Treat any range as a sanity check, not a budget.

The comparison that matters is not between models. It is between quotes, after you price in what each leaves on your side of the table.

Worked example: $0.04 vs. $0.06, priced end to end

You need 100,000 bounding-box labels. Two vendors quote:

• Quote A: $0.04 per label. In the pilot, 92% of labels passed your QA. Rejected labels are re-billed.
• Quote B: $0.06 per label. In the pilot, 99.5% passed. The vendor redoes rejected labels at no cost.

Shared assumptions: checking one label takes 6 seconds, and the person checking costs $95 an hour fully loaded (roughly a $200K total-compensation engineer over 2,080 hours).

The deciding difference is how much you have to review. With 8% of A's labels wrong and no way to know which, you review all of them. With B at 99.5%, a 3% random sample (3,000 labels) is enough to catch a slip.

Quote A's sticker price is 33% lower. Its real cost is 3.3x higher, and 80% of that cost is your own engineers' time.

"We'd just spot-check A." Then review 20% of A's labels instead of all of them. The total drops to $7,484, or $0.075 per label, still more than B. And you have knowingly shipped about 6,400 wrong labels into your training set, because a 20% sample catches only 20% of the errors. That cost shows up later, in model quality, where it's much harder to price.

The math changes if your reviewer is cheap, the task is trivially fast to check, or the low-cost vendor's acceptance is close to the expensive one's. That's why the calculator below lets you change every assumption.

Run your own quotes

Enter two quotes and your team's numbers. The calculator recomputes as you type and shows where each dollar goes: vendor fees, your review time, and rework.

Three inputs move the answer most:

1. Acceptance rate. Get it from a pilot on your data, not from the vendor's website. A pilot of 300 labels tells you whether a vendor is near 99% or nowhere close.
2. Share you must review. Be honest. If you wouldn't ship a vendor's labels unchecked, you're reviewing 100%.
3. Loaded cost of your reviewer. Use the real person. If it's your CTO on a Sunday, price it accordingly.

What else should you price in?

The calculator covers the big three. These smaller items rarely appear on a quote, and each one has blown up a labeling budget somewhere:

• Ramp time. Weeks between signing and the first usable batch are weeks your model doesn't move. Ask for the date of the first production batch, not the start date.
• Guideline churn. Your label definitions will change. Ask whether relabeling after a guideline change is billed, and at what rate.
• Minimums and commitments. Monthly minimums turn a cheap per-label rate into an expensive one in a slow month.
• Platform seats and usage fees. Software-plus-service vendors may bill seats, storage, or usage units on top of labor.
• Integration work. Getting data into the vendor's tool and labels back into your bucket in your schema. Someone on your side builds and maintains it.
• Management time. Weekly syncs, Slack threads, escalations. Count the hours for a month; most teams are surprised.

Six questions to put to any quote

1. What exactly does your accuracy figure measure, and on how large a sample?
2. What happens to labels that fail our QA: redone free, re-billed, or argued about?
3. How many days from kickoff to the first production batch?
4. Do you deliver a QA report with every batch, and what does it contain?
5. What does a guideline change cost us mid-project?
6. Are there minimums, platform fees, or usage charges outside the per-label price?

A vendor that answers all six in writing is easier to compare, and usually easier to work with.

What does Akmon charge?

It would be odd to write this guide and hide our own numbers, so here they are. Prices depend on your schema, volume, and turnaround; these bands are where most programs land.

And our answers to the six questions above:

1. Accuracy: 99%+, guaranteed, measured against your QA.
2. Failed labels: redone at no cost.
3. Kickoff to first batch: live within hours; first production batch within a day.
4. QA report: with every delivery, alongside labels in your format and your bucket.

When we're the wrong choice: very small one-off jobs that your team can finish in an afternoon, or tasks so simple that a model-assisted tool and a quick human pass are enough.

Send us your current quote and pilot acceptance rate, and we'll run this calculator with you, including on the quotes that aren't ours. Book a call.

FAQ

How much does data labeling cost in 2026? Published ranges run from about $0.03 to $1.00 per label, or roughly $4–20 an hour depending on task difficulty. Those ranges hide the bigger cost: your team's review and rework time. Compare quotes on cost per usable label instead.

What is cost per usable label? Everything you spend to get one label that passes your QA, including vendor fees, your review time, rework fees, and re-review time, divided by the labels you needed.

Is per-hour or per-label pricing better? Per-label is easier to budget for large, stable tasks. Per-hour suits complex or changing tasks. For production work where errors are costly, prefer pricing tied to a clearly defined accuracy guarantee with free rework.

How do I measure a vendor's acceptance rate before signing? Run a paid pilot of a few hundred items from your real data, graded by your team against your guidelines. If a vendor gets 300 out of 300 right, you can be 95% confident its error rate is below 1%.

Is in-house labeling cheaper? Sometimes, for small and stable volumes. In-house annotators cost roughly $42K–$90K a year each before management time, tooling, and coverage gaps. Price the hours your engineers spend managing them, too.

Sources

• Averroes: data labeling cost (per-label range; in-house annotator cost)
• AI Superior: AI data annotation cost 2026 (hourly ranges)
• Digital Divide Data: annotation pricing models (per-label, per-hour, outcome-based; cost per usable label)
• CostBench: Scale AI vs. Labelbox pricing (unpublished pricing; usage fees)
• Tian Pan: the eval bottleneck (2026 AI engineer compensation)
• Worked example and calculator: Akmon's own model, formulas shown above. Sample-size claim: one-sided 95% Clopper–Pearson bound.