PUBLISHED PROVIDER LIMITS
Limits keep their provider-specific dimensions.
All API usage is rate limited. Current numeric limits vary and should be read from the account limits page rather than treated as universal constants.
Authentication: OpenAI API key
Primary scope: Organization, Project, Model, And Usage Tier
Measured dimensions: Requests, Tokens, Images, Audio, Or Other Model Specific Units
Profile coverage: Account-, project-, model-, and usage-tier variability plus the common request, token, and project-token response headers. Batch, Realtime, image, audio, and endpoint-specific numeric limits remain account- or model-specific.
Rate limits
| Limit | Value | Period | Scope | Conditions and variability |
|---|---|---|---|---|
| API usage limits | Provider-specific | Provider-defined | Organization Project Model Tier | Varies by model, project, organization, usage tier, and approved limit increases. Retrieve the applicable values from the OpenAI limits settings for the account. |
RESPONSE METADATA
Headers expose the applicable limit state.
X-RateLimit-Limit-RequestsFormat: Integer- Request allowance for the applicable limit.
X-RateLimit-Remaining-RequestsFormat: Integer- Requests remaining for the applicable limit.
X-RateLimit-Reset-RequestsFormat: Duration- Duration until request capacity resets.
X-RateLimit-Limit-TokensFormat: Integer- Token allowance for the applicable limit.
X-RateLimit-Remaining-TokensFormat: Integer- Tokens remaining for the applicable limit.
X-RateLimit-Reset-TokensFormat: Duration- Duration until token capacity resets.
X-RateLimit-Limit-Project-TokensFormat: Integer- Project-scoped token allowance when a project token limit applies.
X-RateLimit-Remaining-Project-TokensFormat: Integer- Project-scoped tokens remaining when a project token limit applies.
X-RateLimit-Reset-Project-TokensFormat: Duration- Duration until project-scoped token capacity resets.
Retry-AfterFormat: Integer Seconds- Minimum delay before retrying a temporary rate-limit or model-overload response, when supplied.
WHEN A LIMIT IS EXCEEDED
Read the response before choosing a delay.
- Rate-limit responses can use HTTP 429.
- Temporary rate limits use HTTP 429; quota, billing, and spend-limit errors can also use 429 and require corrective action. Temporary model overload can use HTTP 503. Inspect the error body before retrying.
RECOMMENDED HANDLING
Protect both the provider and your own queue.
- Read current account-specific values from account settings; this profile publishes no private account quotas.
- Honor Retry-After when supplied for a temporary limit, otherwise use bounded exponential backoff with jitter.
- Do not retry quota, billing, or spend errors that require account action.
- Inspect response headers and error details instead of assuming a universal model limit; avoid replaying consumed streaming output.
VERIFICATION RECORD