Amazon SageMaker Serverless Inference is designed for irregular or unpredictable traffic patterns. It automatically provisions and scales compute resources based on request volume and scales down to zero when idle, making it the most cost-effective option.

Serverless inference supports payloads up to 6 MB and request durations up to 60 seconds, which comfortably meets the stated constraints. Customers are billed only for actual compute usage during inference execution, not for idle capacity.

Asynchronous inference is intended for long-running jobs (up to 1 hour) and large payloads (up to 1 GB). Real-time inference requires always-on instances, increasing cost during idle periods. Batch transform is designed for offline processing.

Therefore, serverless inference is the optimal choice.