Errors, limits, and retries

Error codes tell your application whether to change the request, wait, or ask for user action.

Code	Action
`invalid_api_key`	Check the key. Create another key if it was revoked or expired.
`model_not_found`	Use an exact ID from the model catalog.
`model_unavailable`	Retry later or select another ready model.
`insufficient_balance`	Add credits or use another funded workspace.
`context_length_exceeded`	Reduce the input or output limit.
`unsupported_parameter`	Remove or change the named parameter.
`capacity_exceeded`	Retry with backoff.

Retry rules

Retry 429 and temporary 503 responses. Use exponential backoff, add random delay, and set a retry limit.

Do not retry 401, 404, or parameter errors without changing the request.

A 429 response means that no model slot is available now. The gateway does not queue the request.

Keep the response request or generation ID in your application logs. Do not log prompts, outputs, or API keys with it.