Limits and retries

Key RPM, TPM and concurrency settings, model context/output limits, provider capacity and file quotas apply independently. Current values come from the console, model details and file reference; there is no universal quota for every model.

Distinguish errors

• 402: insufficient personal balance or exhausted project/key budget. Adjust the relevant balance or budget; immediate retries do not help.
• 429: rate or route capacity limits. Reduce concurrency and respect Retry-After when present. The header is not guaranteed on every error.
• 413: request or file is too large. Reduce it before trying again.
• 400/403/404: check parameters, model permissions and exact model ID.

Retry safely

Use bounded exponential backoff with jitter only when retrying is safe. A timeout does not prove the upstream did no work. Creation/generation requests have no general idempotency guarantee and can incur duplicate work or cost. Never silently replay a started stream. See streaming and retries and error codes.

Queue burst traffic and bound Agent turns, tool calls and total spend. Raising a monetary budget does not increase RPM, TPM or provider capacity.