Limits keep a model from consuming more than you intend. There are two levels: the model as a whole, and each member’s access to it.
Where to set them
- Model total — on the model’s Quota tab.
- Per member — in the Members dialog when adding, inviting, or editing a member.
A limit left blank means no limit for that item. An empty budget does not mean zero; it means unbounded.
Kinds of limit
| Limit | Unit | Notes |
|---|---|---|
| Budget | USD | Based on reported request cost, not recomputed from the model’s prices. |
| Tokens | tokens | Counts input and output tokens. |
| Requests | requests | Counts inference requests. |
Each limit has its own period. When a period ends, its counters start again; the console shows the reset time when a period is in use. A model’s limits apply on top of the limits of each member: a member’s budget is not a separate balance, but an additional ceiling. A request must fit within both.
Call protection
The Quota tab also lists concurrency, request rate, and maximum output for the model. These are separate from budgets: they bound how a request is served rather than how much it may cost.
Changing limits
Lowering a limit does not clear what has already been spent, and raising one does not reset a counter. If a budget is already exhausted, raising the budget takes effect immediately; requests are accepted again as soon as the new limit leaves room.
When a limit is reached
- Requests are refused once a limit is reached, with a
429response. - The model card and the detail page show spent against limit; the progress bar turns red at the limit and yellow as it approaches.
- For a periodic limit, the page shows when the period resets.
- If the gateway does not report a cost or usage figure, the console shows it as unknown rather than zero, so a missing number never looks like free usage.
Usage reporting is per model
Each model’s limits are independent, even when two models point at the same upstream server or account. Spending on one does not count against another.