Scale AI pricing

About

Scale AI is the data-engine company behind much of the modern AI stack — it collects, curates, and annotates the training and evaluation data that frontier labs and enterprises use to build models. Founded in 2016 by Alexandr Wang and Lucy Guo, Scale grew from a labeling API for self-driving-car datasets into a vertically integrated data operation that bundles pre-labeling software with a global contributor workforce. Today its product surface spans Scale Data Engine (collection, annotation, RLHF, red-teaming, model evaluation across text, image, video and 3D/LiDAR), the Scale GenAI Platform (enterprise generative-AI applications), the Outlier and Remotasks contributor marketplaces, and Donovan, its public-sector / defense offering.

Commercially, Scale is a labor-arbitrage business: it marks up the work of a reported 240,000+ contractors for a 50%+ gross margin, which is exactly why its pricing is private. Revenue scaled with the generative-AI boom — roughly 870M in 2024 (about a 1.5B annualized run rate by year-end) and an estimated 2B in 2025, with applications (GenAI Platform, Donovan) reportedly contributing 200-300M.

The defining 2025 event was strategic, not a price change. In June 2025, Meta invested about 14.3 billion dollars for a roughly 49% non-voting stake (valuing Scale at 29B) and hired founder-CEO Alexandr Wang to lead its superintelligence effort; Jason Droege became Scale’s CEO. The deal reshaped Scale’s customer base — rivals including Google (which had reportedly spent about 150M in 2024) and OpenAI pulled back over data-confidentiality concerns — pushing Scale further toward enterprise and public-sector contracts.

For current pricing, you book a demo at Scale’s pricing page — there is no public rate card to quote from.

Pricing summary : How Scale AI’s pricing model works

Scale AI is sales-only: the core data business is priced by quote, structured around per labeled task / per annotation / per data unit, and usually packaged as a committed annual data-engine contract with volume discounts. There is no published rate card and no ongoing free tier — only a free trial. The pricing page presents two paths, and both route to “book a demo”:

1. Self-Serve Data Engine — pay-as-you-go via credit card for experimental or research projects, beyond a free trial of the first 1,000 labeling units and first 10,000 images. Per-unit rates are not published on the page.
2. Enterprise — committed contracts spanning Data Engine plus the GenAI Platform, with enterprise SLAs and dedicated customer-operations support. This is where the revenue is, and it is fully sales-quoted.

Underneath, the unit economics are a markup on human labor: Scale monetizes per task and, per third-party reporting, runs a 50%+ gross margin. Third parties cite indicative self-serve figures — roughly 2 cents per image and 6 cents per annotation — and an average contract near 93K per year with complex projects past 400K, but Scale itself publishes none of this.

What makes this different: Scale is one of the largest AI vendors with essentially zero public price transparency for its main product. Where a GPU cloud publishes a per-hour rate, Scale deliberately keeps per-task pricing behind sales — because the price is really a negotiated markup on a labor supply chain it would rather not expose.

Pricing by product

Scale does not publish list prices. The table below describes the buying motion per surface; any cent- or dollar-level figure is a third-party indicative estimate, not an official Scale rate.

Sales motions across products: every customer-facing surface is sales-led and quote-based. The free trial is the only self-serve entry point; there is no published per-task rate card.

Hidden costs : What Scale AI users actually pay

Because Scale prices by negotiated contract, the “hidden costs” are less about line-item add-ons and more about the structure of a quote-based data deal:

The single biggest “cost” is price opacity itself: with no rate card, buyers cannot benchmark a quote without an RFP across vendors, and effective per-unit cost depends heavily on complexity and quality bar (third-party benchmarks for the category span roughly 0.01 to 1.00+ dollars per image, and 0.05 to 3.00 dollars per segmentation mask). The second is vendor concentration risk that the Meta deal made concrete — committing annual volume to a vendor partly owned by a competitor is now a real procurement consideration.

Want to estimate your own data-labeling bill? Use the Scale AI pricing calculator to model costs against indicative per-task assumptions.

Pricing evolution : Scale AI pricing history and changes

Cadence

Tracked range: 2023–present. Pricing itself stayed quote-based throughout; the material changes were strategic (ownership, customer base), not rate-card moves.

Notable changes

• 2023 — Per-labeled-task data-engine model scales with the generative-AI training-data boom; frontier labs buy RLHF and annotation at volume. No public rate card.
• 2024 — Enterprise committed contracts dominate revenue (~870M). Self-serve Data Engine offers a free trial but routes paid usage to sales.
• June 2025 — Meta takes a roughly 49% non-voting stake for about 14.3B (29B valuation); Alexandr Wang departs for Meta, Jason Droege becomes CEO. Customers including Google and OpenAI reportedly pull back, accelerating the shift toward enterprise apps and public-sector Donovan.

The throughline: Scale has barely changed its pricing mechanic in years — it is still per-task, contract-based, sales-quoted. What changed in 2025 was who buys and who owns the vendor, which matters more for a quote-based business than any list-price tweak.

What’s unique : Scale AI’s distinctive pricing mechanics

1. A market leader with no public rate card. Unlike most of the AI-infra corpus, Scale publishes essentially nothing — both the self-serve and enterprise paths route to “book a demo.” Pricing is a negotiated markup on labor, kept private by design.

2. The unit is human labor, not compute. Scale meters per labeled task / per annotation and earns a reported 50%+ gross margin by marking up a 240,000+ contributor workforce. The “value metric” is annotated data units, not GPU-hours or tokens.

3. Ownership became a pricing variable. After Meta’s ~49% stake, a procurement decision about Scale is also a decision about feeding a competitor — a rare case where the cap table, not the rate card, drives buyer behavior.

Strengths & weaknesses

Billing UX : Scale AI billing controls and transparency

• Billing controls — Self-serve Data Engine bills pay-as-you-go via credit card after the free trial; enterprise runs on committed annual contracts with invoicing and volume discounts negotiated by sales.
• Usage visibility — The Data Engine console tracks labeling units and project spend, but there is no published rate to reconcile against — effective per-task cost is set in the contract.
• Payment options — Credit card for self-serve; sales-led contracts, POs and invoicing for enterprise and public-sector (Donovan) deals.

Strategic wins : Why Scale AI’s pricing decisions worked

1. Pricing the data, not the model

By charging per labeled task instead of per seat or per model, Scale captured value from the single scarcest input in AI — high-quality training data — and turned a BPO motion into a high-margin software-adjacent business. See how AI companies structure pricing.

2. Opacity as leverage

Keeping per-task rates private let Scale price each frontier-lab deal to willingness-to-pay rather than a public anchor, protecting its labor markup. Related: outcome-based pricing trends.

3. Layering committed contracts over self-serve

A free trial funnels experiments into committed annual data-engine deals, converting spiky project demand into forecastable enterprise revenue. See choosing the right usage metric.

Areas to improve : Gaps in Scale AI’s pricing approach

1. No way to estimate before talking to sales

With no published per-task rate, a buyer cannot ballpark a budget without an RFP. Even an indicative range or a public self-serve rate would reduce friction at the top of the funnel. See bill shock and cost unpredictability.

2. Complexity-driven cost is unpredictable

Effective per-unit cost swings widely with annotation type and quality bar, and the contract structure makes that hard to forecast — clearer complexity tiers would help buyers plan.

3. Ownership conflict needs a pricing answer

Post-Meta, Scale needs commercial guardrails (data isolation guarantees, neutral-vendor assurances) priced into deals to win back labs that pulled back — otherwise quote-based flexibility cannot offset the trust gap.

Key takeaways

1. Scale AI is sales-only, per-task pricing — no public rate card, contract-based per labeled task / annotation / data unit, with committed annual data-engine deals. For the underlying model, see the introduction to usage-based pricing.
2. The meter is human labor, not compute — Scale marks up a 240,000+ contributor workforce for a reported 50%+ gross margin, which is why pricing stays private.
3. A free trial is the only self-serve door (first 1,000 labeling units, first 10,000 images); everything past it is quoted.
4. The big 2025 change was strategic, not a price move — Meta’s ~49% / ~14.3B stake and the CEO swap reshaped Scale’s customer base more than any rate-card tweak.
5. Opacity is the model’s signature and its weakness — it protects margin but blocks buyer benchmarking and amplified the post-Meta trust problem.

UBP implications

1. Usage pricing can be fully private and still scale. Scale proves a per-unit, usage-shaped model doesn’t require a public rate card — but only if you own a defensible, hard-to-benchmark input.
2. The value metric should track the scarce input. Pricing per annotated data unit aligns cost with the thing customers actually need (quality data), not a proxy like seats.
3. Ownership and trust are pricing inputs for data vendors. When a buyer’s data feeds a vendor partly owned by a rival, no discount fully closes the gap — neutrality has to be built into the offer, not just the price.

Sources

• Scale AI pricing page — sales-quoted, free trial only (accessed 2026-06-15)
• Scale Data Engine — no prices disclosed (accessed 2026-06-15)
• Sacra — Scale AI revenue, valuation & model — indicative per-unit figures, revenue, Meta deal (accessed 2026-06-15)
• CheckThat.ai — Scale AI pricing (third-party) — indicative per-unit and contract figures (accessed 2026-06-15)
• CNN Business — Scale AI making money under Meta — CEO Droege on revenue mix (accessed 2026-06-15)
• AInvest — Meta acquires 49% stake in Scale for $14.3B (accessed 2026-06-15)

Bottom line

Scale AI is the rare market leader that publishes almost no pricing: its data engine and GenAI platform are sold by enterprise quote, structured around per-labeled-task / per-annotation rates and committed annual contracts, with only a free trial (first 1,000 labeling units, first 10,000 images) as a self-serve door. The unit is human labor — Scale marks up a 240,000+ contributor workforce for a reported 50%+ gross margin — which is exactly why the rate card stays private. The defining 2025 event was strategic, not a price change: Meta’s roughly 14.3B investment for a ~49% non-voting stake pulled founder-CEO Alexandr Wang to Meta and reshaped Scale’s customer base as rivals like Google and OpenAI pulled back. Browse the pricing blueprint for more fully-researched company profiles.

Want to compare Scale AI against other Infrastructure, Compute & MLOps companies? Browse the pricing blueprint.