Skip to main content
Before starting a batch, check sandbox concurrency limits to see how much room remains for new work. A funded account can still receive an admission refusal:
The native operation is GET /v1/limits.

​
What the response contains

The public pricing catalog also publishes admission tiers and hard batch maxima. GET /v1/pricing is the live source for that catalog and rates. Per-request enforcement remains authoritative.

​
Other limits

Fleet requests have a hard maximum of 10,000 machines; swarms 256; forks 20. Your effective policy, aggregate quota, funds, and capacity can allow fewer. Large fleets use asynchronous status polling; see Create fleets. The estimator accepts up to 256 machines per bounded estimate. This is an estimator input bound, separate from fleet creation limits. Trial accounts have additional restrictions. Machine sizes define resource capacities; availability depends on placement and live capacity.

​
Errors and what to do

A limit rejection includes error_code, a named limit, maximum, usage, reset_at when applicable, and next_action. Start windows use fixed UTC minute, hour, and day boundaries. Reset times are Unix seconds. Honor Retry-After when present. Resource exhaustion may have no clock-based reset: pause or delete work, reduce the request, or wait for capacity to be released. Paid bursting can extend start allowances with consent. It does not bypass other admission checks. See Paid extra starts.