Limits and errors

Per key per minute: chat 60, images 10, vision 30, reads 120. Errors use the OpenAI-compatible shape {"error": {"message", "type", "code"}}; common codes: invalid_api_key (401), insufficient_balance / key_spend_limit_reached (402), rate_limited (429), ai_busy (503). Send Idempotency-Key on POSTs so a retry never charges twice.