AI Training Data Pricing Guide 2026: What You Should Actually Pay

Why pricing is so opaque in AI training data

AI training data pricing resists comparison more than almost any other B2B service category. Vendors quote per-image, per-hour, per-annotation, per-task, or per-deliverable -- sometimes all four in the same RFP response. The unit price says almost nothing about total cost without knowing what counts as a unit, what QA is included, what rework policy applies, and whether project management is bundled.

This opacity is not accidental. Many vendors quote a low per-unit rate and recover margin through project management fees, correction cycles, and minimum-spend thresholds. A buyer who awarded on unit price alone often finds that the final invoice is 40 to 80% above the quoted rate.

This guide gives you a cost framework by modality, the main cost drivers, and the questions to ask before you sign anything.

Annotation pricing by modality (2026 benchmarks)

Image annotation rates vary by complexity. Simple bounding boxes on clean images run $0.05 to $0.15 per image from a managed offshore provider. Semantic segmentation on the same image runs $0.50 to $2.00 depending on the number of classes and polygon density. LiDAR 3D cuboid annotation runs $2 to $8 per frame for a single sensor, rising to $5 to $20 per fused LiDAR-camera frame.

Text annotation rates depend on task type. Binary classification (spam, not spam) runs $0.02 to $0.08 per item. Named entity recognition on short documents runs $0.10 to $0.40 per document. Summarization quality evaluation by domain-expert annotators runs $0.50 to $3.00 per item, and complex reasoning evaluation by PhD-level reviewers can reach $5 to $15 per item.

RLHF preference annotation is the most variable segment. Generic instruction-following preference pairs run $0.50 to $2.00 per pair from a general workforce. Technical domain preference annotation -- code, math, medicine -- by domain experts runs $5 to $25 per pair and sometimes higher for specialist knowledge. Red-teaming (adversarial prompting) by experienced practitioners runs $8 to $30 per hour of structured testing.

Speech data runs $0.10 to $0.50 per minute for scripted read-aloud collection in common languages. Spontaneous conversation collection is 2 to 3 times higher. Rare language and dialect coverage commands a premium of 50 to 150% above equivalent common-language rates. Transcription of existing audio runs $0.05 to $0.15 per minute for clean audio, rising to $0.20 to $0.40 per minute for accented or noisy speech.

Field and physical-world data collection pricing

Field data collection -- driving data, physical object capture, environment mapping -- is quoted by the day or by the capture unit rather than per annotation. A field capture team (two to four people with equipment) typically runs $400 to $1,200 per day in developed markets and $150 to $500 per day in lower-cost regions. Equipment rental, travel, and permitting add to the base rate.

Drone survey and aerial capture runs $800 to $3,000 per flight day depending on equipment complexity and location. Data quality, including GPS accuracy logs and metadata, is bundled into higher-tier quotes; budget vendors often deliver raw footage without the structured capture logs that downstream annotation requires.

For IoT and sensor data, expect a setup and integration fee of $2,000 to $10,000 followed by a data delivery rate that depends on volume. Physical product capture for e-commerce (360-degree imagery, dimensional measurement) runs $1.50 to $8.00 per SKU at scale.

What drives cost above the quoted unit rate

Rework is the largest hidden cost. If a batch fails your accuracy threshold -- whether at 85%, 90%, or 95% -- the vendor must re-annotate. If your contract does not clearly specify whether rework is included in the unit rate or billed separately, you may face a second invoice for work that should have been right the first time. A vendor with a three-tier QA cascade catches errors internally before delivery; one that passes quality back to the client as a revision cycle bills for it.

Ontology and calibration overhead adds cost on complex projects. Before a vendor can annotate your data accurately, they need to understand your taxonomy, calibrate on your edge cases, and pass a gold-set evaluation. This can take 20 to 80 hours of setup time. Some vendors bill setup separately ($150 to $300 per hour); others build it into the first batch's effective rate. Either way, it is real cost that does not appear in a per-unit quote.

Rush pricing is real. Most vendors can deliver at standard turnaround (48 to 72 hours for moderate volumes). For same-day or next-day delivery, expect a 25 to 75% premium. Weekend work typically costs 50% more than weekday rates. If your ML training pipeline runs on tight feedback cycles, time-of-delivery costs are structural, not exceptional.

Project management and communication overhead is often unbundled. A fully managed vendor -- one who does task design, annotator training, daily QA reporting, and delivery formatting without requiring significant input from your team -- charges more per unit than a platform that puts those tasks back on you. The right comparison is total team-hours cost, not vendor invoice alone.

Where buyers overpay: the three most common errors

Overpaying for domestic rates when offshore delivery is equally accurate is the most common error. US-based annotation workers charge $20 to $45 per hour for general tasks and $60 to $150 per hour for specialized RLHF. Managed offshore providers with equivalent quality architecture deliver at $4 to $12 per hour equivalent for the same work. For high-volume programs, the gap compounds quickly. Corpshore AI typically delivers at 50 to 70% below US-domestic rates; see current indicative costs at https://corpshore.ai/pricing.

Paying platform fees on top of labor is the second common error. Several major vendors are annotation platforms that take a percentage cut of every annotation task passed through their system, regardless of who does the work. If you supply your own data, your own guidelines, and use the platform for workforce sourcing only, the platform fee can represent 30 to 50% of total cost for something that a managed operator would bundle into the delivery rate.

Accepting per-unit quotes without a rework policy in writing is the third error. A quote of $0.08 per image sounds clean until 15% of the first batch fails QA and you are invoiced for the re-annotation at full rate. Require a written SLA that specifies accuracy thresholds and what happens when they are missed.

How to build a realistic budget

Start with a task decomposition, not a volume estimate. For a 1 million item annotation project, the budget question is not 'how much per item' but 'what is the distribution of task types, complexity levels, and quality tiers.' A million items that are 80% easy classification and 20% complex polygon segmentation costs very differently than a million items with even distribution.

Build in 20% headroom for rework, calibration, and scope changes. AI training data projects almost always evolve as the model team's understanding of the data matures. Ontology changes mid-project, additional classes get added, sample rates get adjusted. A budget that has no flexibility will create friction at exactly the moment the project is accelerating.

Get a reference from at least one comparable project. Ask the vendor for an example of a similar project -- modality, volume, language, complexity -- and what the final cost per unit came out to after rework, setup, and QA. The gap between quoted rate and actual cost on a comparable project is the most honest signal of a vendor's cost predictability.

Frequently asked questions

How much does AI data annotation cost in 2026?

AI annotation costs vary widely by modality and complexity. Simple image classification runs $0.05 to $0.15 per image; semantic segmentation runs $0.50 to $2.00; RLHF preference pairs run $0.50 to $25 per pair depending on domain expertise required. Managed offshore providers typically deliver at 50 to 70% below US-domestic rates.

How much does RLHF annotation cost?

RLHF preference annotation costs $0.50 to $2.00 per pair for general instruction-following tasks and $5 to $25 per pair for technical domains (code, math, medicine) requiring domain-expert annotators. Red-teaming runs $8 to $30 per hour of structured testing.

Why is AI training data pricing so opaque?

Vendors quote in different units (per item, per hour, per deliverable) and often exclude rework, setup, and project management from their headline rate. The final cost is typically 40 to 80% above a naive per-unit estimate when these factors are included.

What is the cost advantage of offshore AI annotation?

Managed offshore annotation providers with equivalent quality architecture typically deliver at 50 to 70% below US-domestic rates. The key caveat is 'equivalent quality architecture' -- a low offshore rate with poor quality outcomes costs more after rework than a higher domestic rate with consistent accuracy.

How much does speech data collection cost?

Scripted speech collection runs $0.10 to $0.50 per minute for common languages. Spontaneous conversation collection is 2 to 3 times higher. Rare language and dialect coverage commands a 50 to 150% premium. Transcription of existing audio runs $0.05 to $0.40 per minute depending on audio quality.

How do I get a cost estimate for my AI training data project?

Start with a task decomposition: break the project into task types, complexity levels, and quality tiers. Then get a per-unit quote with a written rework and SLA policy, and ask the vendor for an actual final cost on a comparable past project. Corpshore AI provides indicative scoping at corpshore.ai/pricing or via a pilot scoping call.