A company must deploy a custom-trained classification model on AWS to provide near real-time, low-latency predictions and to handle varying request volumes. Which deployment option meets these requirements?

Deploying the model as an Amazon SageMaker real-time endpoint with auto scaling directly addresses the requirements. SageMaker real-time endpoints are designed for low-latency, near real-time predictions and can handle varying request volumes efficiently through auto scaling. Batch transform jobs are suitable for offline, high-throughput inference on large datasets, not near real-time. Using API Gateway with S3 for model files would require significant custom development for inference logic, model loading, and scaling, making it less efficient and more complex than SageMaker. Launching EC2 instances with Deep Learning AMIs behind an Application Load Balancer offers flexibility but requires manual setup and management of the inference server, scaling, and monitoring, which SageMaker endpoints provide out-of-the-box.