Annotation quotes come in two main shapes. One vendor says “two cents a box”. Another says “six dollars an hour”. Both can be fair, and both can go badly wrong for the buyer if the shape does not match the work.
What does each model actually charge for?
Per task (or per object, per item, per entity) charges a set amount for each unit of output: each bounding box, each keypoint, each tagged entity, each transcribed minute. If some items turn out slow, that is the vendor’s problem.
Per hour charges for annotator time, whatever gets done in it. You carry the risk of slow work, and you gain when the work turns out to be quick.
Few vendors publish either. Of the annotation vendors we compared on November 7, 2026, Label Your Data is the only managed service with public rates, and it shows both shapes on one page:
| Published rate (Label Your Data) | Unit |
|---|---|
| From $0.015 | per keypoint |
| From $0.02 | per bounding box |
| From $0.02 | per NLP entity |
| Typically $6 | per annotator hour, for hourly projects |
Platforms price differently again. Segments.ai sells labeling usage by the year, starting at 3,600 hours, which it says is equivalent to $2.67 per hour. HumanSignal bundles a dozen hosted Label Studio seats for $99 monthly. Those are tool costs; the people doing the labeling are on top.
When is per-task pricing the better deal?
Per task works best when three things are true:
- The items are uniform. Product photos with one or two objects each, or short texts with a fixed label set.
- The guidelines are settled. Every rule change mid-project becomes a price discussion.
- You can count the units. You know roughly how many boxes or entities each item holds.
It gives you a firm budget and pushes the vendor to work efficiently. The risk is density: at two cents a box, a street scene with 60 objects costs $1.20 while a product shot costs two cents. If your data mixes both, ask for a per-image price band, or a cap.
Watch for a quieter risk too. Paid per unit, a crew is rewarded for speed, and quality can slide on hard items. A per-task contract needs a firm quality clause with a gold set and an audit.
When is per-hour pricing the better deal?
Per hour suits work where output per hour is hard to predict:
- Early or exploratory projects, where the label schema is still changing.
- Complex judgment, such as rating long LLM answers, reviewing legal clauses or grading code.
- Expert work, where the annotator’s time is the scarce thing and the output varies item by item.
It keeps the vendor from cutting corners to hit a unit rate, and it makes guideline changes cheap. The risk is that slow work costs you money. Protect yourself with a target pace agreed after the pilot, weekly reporting of items per hour, and a cap on total hours per delivery.
How do you compare a per-hour quote with a per-task quote?
Convert one into the other with the vendor’s own pace figure:
Cost per item = hourly rate ÷ items per hour × (1 + review and rework share)
A worked example, with numbers you should replace with your own:
| Input | Value |
|---|---|
| Hourly rate | $6 (Label Your Data’s published typical rate) |
| Items per hour | 60 boxes |
| Review and rework | 20% |
| Cost per box | $6 ÷ 60 × 1.2 = $0.12 |
That is well above the published “from $0.02” box rate, which tells you one of two things: the per-task rate assumes a faster pace or less review, or the “from” price applies only to simple work. Either way, it is the question to ask. Our annotation pricing page has a converter that does this sum.
What else changes the real cost?
Headline rates leave out a lot. Before you compare, ask each vendor:
- Is quality review included, and how many review passes?
- Is rework on rejected items free, and for how long?
- Are project management and tooling included or extra?
- Is there a setup fee, a minimum order, or a minimum term?
- Do prices change after the pilot, or at volume thresholds?
A vendor with a higher unit price and everything included can come out cheaper than a low rate with paid rework.
Can you mix the two models?
Yes, and many projects should. A common pattern is to pay hourly for the pilot and calibration, while the guidelines are still moving, then switch to per-task pricing once pace and quality have settled. Expert review or edge-case handling can stay hourly while the bulk of simple items moves to a unit rate.
Write the switch point into the contract: for example, “per-task rates apply from the first delivery after agreement on the pilot reaches the target”. That keeps both sides honest about when the work became predictable.
Which should you ask for?
Ask for both. Send every vendor the same sample and request a per-task price, an hourly price and their expected items per hour. If the two prices do not roughly agree once converted, ask why. The answer usually tells you more about how the vendor works than either number does.
When you are ready to compare vendors side by side, the ranking of the best data annotation companies shows which ones publish prices at all.
Pricing is the one part of an annotation contract you can check with a calculator, so check it.