Skip to main content
Budgets can be set at six levels: Workspace, Project, Identity, API Key, Provider, and Model. When several budgets apply to the same request, every one of them is enforced independently and cumulatively. There is no precedence: a budget at one level never overrides a budget at another level, and a request must satisfy every budget that matches it. Set up and manage budgets under Settings > Organization > Budgets. See Budgets for creating budgets, configuring limits, and setting up alerts.

​
Budget levels

Only one budget is allowed per target, so a request can match at most one scoped budget per level, up to six in total. Dynamic budgets created through the API with a custom matching expression can apply in addition. When a budget is exhausted, only requests that match it are rejected.

​
How multiple budgets interact

  • Every applicable budget is enforced. A request must pass every budget that matches it. No budget overrides another.
  • Each budget tracks its own usage. A request counts toward every budget that applies to it, at the same time.
  • The most restrictive budget binds. The effective cap for a dimension is the most restrictive applicable budget.
  • Requests per minute are enforced per budget, each with its own rolling 60-second window.
  • Expired budgets stop enforcing. After the expiration date a budget no longer blocks requests, even though it stays in the list.

​
Worked examples

The workspace has a $1,000/month budget. The Production project has a $300/month budget.
  • Requests to Production count toward both budgets; requests to other projects count only toward the workspace budget.
  • When Production reaches $300, requests to Production are rejected until the monthly reset. Other projects keep working until the workspace budget is exhausted at $1,000.
The Support project has a $500/month budget. The identity customer-acme has a $50/month budget.
  • The identity budget applies across all projects: the $50/month cap is enforced whenever customer-acme makes a request, regardless of project.
  • Requests from customer-acme count toward both budgets. When the identity budget is exhausted, those requests are rejected even though the project budget has room; other identities keep working.
The workspace has a $2,000/month budget. The model anthropic/claude-sonnet-5 has an $800/month budget.
  • When the model budget is exhausted, requests to that model are rejected; other models keep working.
  • When total workspace spend reaches $2,000, all requests are rejected until the monthly reset, regardless of model.

​
When a limit is exceeded

A request that matches an exhausted budget is rejected with 429 Too Many Requests. The response identifies the budget and the exceeded dimension (cost, tokens, or requests); the response headers report the limit, remaining usage, and seconds until reset. Only requests that match the exhausted budget are blocked. For budgets with a One-time reset period, there is no automatic reset: matching requests stay blocked until an admin raises the limit, deletes the budget, or its expiration date passes.

​
Reset periods

Each budget has its own reset period and its own counters. Budgets at different levels can use different periods: Daily, Weekly, Monthly, Yearly, or One-time (no reset). See Budgets for the reset schedule. Requests per minute always use a rolling 60-second window, regardless of the reset period.

​
See also

Budgets

Create, edit, and monitor budgets and their alerts.

API Keys

Create and manage AI Gateway API keys with optional spending limits.

Identities

Track user interactions and usage per identity.