Errors and limits

An error is an error object; there is never a partial verdict.

{
  "error": {
    "message": "Input exceeds the 4194304-byte limit for this key.",
    "type": "invalid_request_error",
    "code": "input_too_large"
  }
}

Status	code	When	What to do
400	invalid_request, invalid_json, invalid_utf8, invalid_query, empty_input, invalid_role, invalid_kind, unsupported_content, too_many_messages, model_not_found, invalid_body	The request is wrong.	Fix the request; the message says what.
401	invalid_api_key	The key is missing, unknown, expired or revoked.	Check the Authorization header and the key.
403	access_restricted	The key is valid, but access is restricted for it or its organization.	Ask your organization’s administrator.
408	upload_timeout	The body stalled or took too long to arrive.	Retry.
413	input_too_large, body_too_large	Over the input limit, or a JSON body larger than its limit.	Send less, or, on /v1/scan, send it as text/plain.
415	unsupported_media_type	The body is neither JSON nor, on /v1/scan, plain text.	Set Content-Type.
422	input_unprocessable	This input can never be scanned as sent.	Retrying will not help.
429	rate_limit_exceeded, scanner_busy	The key’s rate limit, or the service is at capacity.	Wait for Retry-After seconds, then retry.
503	scanner_unavailable	The service cannot scan for this key right now.	Retry with backoff.
504	scan_timeout	The scan ran out of time.	Send less, or split the input.
500	scan_failed	Anything else. No verdict was produced.	Retry with backoff.

Nothing about a request is judged before its key: a request without a valid key gets 401, unless the service itself is unavailable (503).

Rate limits and capacity

A key can have a requests-per-minute limit, set when it is created. It is applied by each server of the service, so a burst can slightly exceed it. When the service is busy, a request for a short input waits up to about two seconds for capacity before it is refused with 429 scanner_busy; a large input waits for capacity as it is scanned. Both 429 answers carry a Retry-After header in whole seconds: for rate_limit_exceeded it is when the key has capacity again, for scanner_busy a short pause before retrying.

Time

A one-line prompt is answered in tens of milliseconds; a megabyte takes tens of seconds. A response that is not streamed must arrive within about two minutes, or the connection is closed on the way: stream anything larger than the model’s context length. Closing the connection cancels the scan.