Bedrock's value is breadth behind one AWS control plane: Amazon's Nova models alongside Anthropic, OpenAI, Meta, Mistral, Cohere, DeepSeek, Qwen and others, all governed by IAM, logged to CloudWatch and billed to your AWS account. That breadth comes with choices that trip teams up: two inference endpoints (bedrock-runtime for Converse and InvokeModel, and bedrock-mantle for OpenAI-compatible and Anthropic Messages APIs, with AWS guidance on which to use for new applications having shifted during 2026), model-specific feature support, cross-Region inference profiles, and four service tiers (Reserved, Priority, Standard and Flex) that trade latency against price.
This course teaches you to make those choices deliberately. You will call models with the Converse API so one request shape works across providers, add tool use and streaming, and plan throughput with inference profiles and tiers. You will cut cost with prompt caching, batch inference, Intelligent Prompt Routing and distillation. You will build retrieval on Knowledge Bases, enforce policy with Guardrails, and evaluate both models and RAG pipelines. The agent module compares classic Bedrock Agents with AgentCore's Runtime, Gateway, Memory, Identity and Policy services. The course closes with custom model import, logging and cost allocation, and a hypothetical reference architecture for a governed internal assistant. Facts are as of October 2026 and checked against the Amazon Bedrock User Guide.