Scale AI

What Scale AI pricing actually meters

A task, and the docs are precise about it for once.

Scale's API reference defines a task as an individual unit of work, with a one-to-one mapping between a task and the data to be labeled. One image, one video, one piece of text, one task, each with its own ID. The GenAI Data Engine docs say the same thing in the same words. Tasks group into batches, batches sit inside projects, and a project binds to one annotation use case.

Annotations are the second leg. When Scale still published rates, every product charged a base rate per item plus a rate per label on it, so the bill moved with scene density rather than file count. Nothing says that split has gone away.

Hours never appear. No contributor-hour unit exists in the customer-facing API, which matters because buyers assume a labeling vendor bills by the hour. The one exception is the AWS listing for Scale GenAI Platform, which adds a dimension called Scale Token Units and then declines to define one. That's the disclosure problem in miniature: Scale names the unit and withholds its size.

What happens when you hit the limit

Nothing public says. Scale documents technical limits and skips commercial ones entirely.

The published ceilings are engineering ones: 30 task-creation requests per second, 2,500 attributes, 10 KB of task metadata, a 120 MB attachment cap, unlimited tasks per batch. Each carries the same footnote, which is to contact support if you want it raised.

On money there's nothing. No spend cap, no throttle, no auto-charge behavior, no word on what happens when a committed volume runs out mid-project. Marketplace buyers get one warning, that entitlements expire when a contract ends without renewal. The rest sits in an order form you see after a sales call.