Different kinds of limits
| Limit | Controls | Where to check |
|---|---|---|
| RPM/TPM, concurrency, window allowance | Request frequency and capacity | Actual responses, account limits and applicable Retry-After |
| Context window | Context handled in one model request | Selected profile context_window |
| Output budget | Maximum output tokens in this request | max_output_tokens and request max_tokens |
| Per-request spending cap | Customer liability for a V2 request | acosmi_options.max_credits |
| Key monthly budget | Cumulative usage within a budget scope | Console Key budgets |
When no public account RPM/TPM value is provided, do not substitute the vendor’s limits. Context length is not a rate limit, and token limits are not spending caps.
Output parameters
Current public text profiles declare max_tokens. Send a positive integer within the model’s lower cap. The prepaid text source contract has an overall hard maximum of 1,048,576, but individual models may be far lower. Some legacy paths fall back to a default budget for omitted/zero values, so examples set it explicitly.
Public prepaid validation and some entry points differ on max_completion_tokens. Until endpoint support is confirmed, use the declared max_tokens; do not send both. Current prepaid text supports n=1 only.
When applicable, thinking.budget_tokens is a finite nonnegative integer no greater than the effective output budget, not a numeric string. Reservation estimates must satisfy the existing exact-integer boundaries; clients cannot invent the server’s billable token estimate.
Request size
The text route source limit is 32 MiB of JSON. The effective deployed limit is the minimum of server, proxy and upstream limits and may be lower. Media and other APIs have their own limits.
429 or timeout
Respect applicable Retry-After, reduce concurrency or request size. For uncertain paid requests, inspect My requests before retrying. See error handling.