Cost of training data estimator

Multiply your volume by a published rate to get a first figure for the cost of training data. Labeling starts at $0.015 to $0.02 per object at the one vendor that publishes per-object rates, licensed video has been reported at $1 to $4 a minute, and an RL task with its grader mostly costs $200 to $2,000.

Rates last verified 7 November 2026. Published starting prices only.

What will your training set cost?

What are you buying?

How many objects?

Share re-done after review, % (your assumption)

Estimated cost

Turn on JavaScript to calculate

Cost = quantity × published rate × (1 + share re-done)

Rate source: Label Your Data pricing page, "from" price (seen 7 November 2026).

Formula: quantity × published unit rate × (1 + your rework share), rates as listed in the table below. Not a quote: vendors price real projects after scoping, and quotes often come in above starting rates. Platform fees, taxes, your own review time and project management are not included.

Which published rates does it use?

The same figures in plain form, with a worked example at the default volume and 10% rework.

Unit	Published rate	Example volume	Example cost	Source
Bounding boxes (per object)	from $0.02	50,000 objects	from $1,100	Label Your Data pricing page, "from" price
Keypoints (per object)	from $0.015	50,000 objects	from $825	Label Your Data pricing page, "from" price
Text entities, NLP (per entity)	from $0.02	100,000 entities	from $2,200	Label Your Data pricing page, "from" price
Managed annotator hours	from $6	500 hours	from $3,300	Label Your Data pricing page, "typically $6 per annotator hour"
Human participant hours (studies, preference ratings)	$8 to $12	300 hours	$2,640 to $3,960	Prolific help center: $8 floor, $12 suggested; fee extra, rate unstated
Licensed video footage (per minute)	$1 to $4	60,000 minutes	$66,000 to $264,000	Bloomberg, 10 Jan 2025: what AI companies were paying
RL tasks with graders (per task)	$200 to $2,000	100 tasks	$22,000 to $220,000	Epoch AI FAQ, 12 Jan 2026, from market interviews

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

What the estimate leaves out

Published rates describe the cheapest version of a task. Real projects add guideline writing, a pilot, quality review, specialist annotators, security requirements and project management. Platform software, if you run your own annotators, is a separate line: at Segments.ai the smallest plan commits you to 3,600 hours yearly at an effective $2.67 each, so roughly $9,612 before a single label is drawn.

Licensed data rarely has a per-unit price at all. The video figure is a reported market rate, not a vendor's list price, and most licenses are negotiated on size, uniqueness and term. See how each pricing model behaves.

Sources behind this section

About the estimator

Is the estimate a quote?

No. It multiplies published starting rates by your volume. Vendors price real projects after scoping, and quotes are often higher than starting rates once quality targets, expertise and security requirements are added.

Why is there sometimes no upper figure?

Some companies publish only a "from" price. We show that floor and do not invent a ceiling.

What does the rework share do?

It adds the cost of items that fail review and are done again. It is your assumption, not a published figure; a pilot is the way to measure it.

Where is expert RLHF data in the estimator?

Expert networks do not publish buyer rates, so there is nothing verified to multiply. RL tasks are included because Epoch AI published a range from market interviews.

Why does the exclusive box multiply RL tasks by 4 to 5?

That multiple comes from the people Epoch AI interviewed about exclusive versus shared RL tasks. It is applied to the RL task line and nothing else.