Amazon Bedrock is a fully managed service that makes foundation models (FMs) from leading AI companies available through a unified API. You don't train models, manage GPUs, or handle infrastructure — you simply call an API and get intelligent responses.

Bedrock is the fastest way to add generative AI to your applications. It abstracts away all the complexity of hosting large language models — you focus on your application logic and prompts.

Feature	Amazon Bedrock	Amazon SageMaker
Purpose	Use pre-trained foundation models	Train and deploy custom ML models
Infrastructure	Serverless — fully managed	Managed instances — you choose type/size
ML expertise	Not required — prompt engineering	Required — data science skills
Customization	Fine-tuning, RAG, prompt engineering	Full control — any algorithm, any data
Cost model	Per token (input + output)	Per hour (instance runtime)
Best for	Generative AI apps (text, image, code)	Custom ML (fraud, forecasting, recommendation)

Bedrock provides access to foundation models from multiple providers — choose based on capability, speed, and cost:

Provider	Model	Strengths	Best For
Anthropic	Claude 3.5 Sonnet, Claude 3 Opus/Haiku	Reasoning, instruction following, safety	Complex tasks, analysis, coding, long context
Amazon	Titan Text, Titan Embeddings, Titan Image	AWS-native, cost-effective, embeddings	Embeddings for RAG, simple generation
Meta	Llama 3.1 (8B, 70B, 405B)	Open-weight, strong performance/cost	General-purpose, customization-friendly
Mistral AI	Mistral Large, Mixtral	Fast, multilingual, code	Low-latency apps, European compliance
Cohere	Command R, Embed	Enterprise search, RAG-optimized	Enterprise search, multilingual RAG
Stability AI	Stable Diffusion XL	Image generation	Marketing visuals, product images

Bedrock offers two approaches to customize models for your domain:

Method	How It Works	When to Use
Fine-Tuning	Train on your labeled data (prompt/completion pairs)	Domain-specific tone, format, or knowledge
Continued Pre-Training	Train on unlabeled domain text (raw documents)	Teach model new domain vocabulary/concepts

The fundamental Bedrock operation — send a prompt, get a response:

• InvokeModel — synchronous, single response (most common)
• InvokeModelWithResponseStream — streaming tokens as they generate (chat UIs)
• Converse API — unified multi-turn conversation API across all models

Feature	On-Demand	Provisioned Throughput
Pricing	Per token (input + output)	Per hour (reserved capacity)
Throttling	Shared limits — can be throttled	Guaranteed — your dedicated capacity
Latency	Variable under load	Consistent — no noisy neighbors
Best for	Dev, variable traffic, experimentation	Production with SLAs, high-volume

Guardrails apply safety controls across any model in Bedrock — a centralized policy layer for responsible AI:

Retrieval-Augmented Generation (RAG) is the technique of grounding AI responses in your actual data. Instead of relying solely on the model's training data, you retrieve relevant documents and include them in the prompt.

Bedrock Knowledge Bases is a fully managed RAG service — connect your data sources, Bedrock handles chunking, embedding, storage, and retrieval automatically.

Bedrock Agents are autonomous AI systems that can reason, plan, and execute multi-step tasks:

• Break complex tasks into steps
• Call external APIs (action groups) to get data or perform actions
• Query knowledge bases for information
• Maintain conversation state across turns
• Handle errors and retry with different approaches

Bedrock uses token-based pricing — you pay for what you use:

Model	Input (per 1K tokens)	Output (per 1K tokens)	Notes
Claude 3 Haiku	$0.00025	$0.00125	Cheapest — high volume tasks
Claude 3.5 Sonnet	$0.003	$0.015	Best value for production
Claude 3 Opus	$0.015	$0.075	Premium — complex reasoning
Llama 3.1 70B	$0.00099	$0.00099	Open model, good price
Titan Text Express	$0.0002	$0.0006	AWS native, lowest cost
Titan Embeddings	$0.0001	—	For RAG vector generation

Full enterprise architecture with private data grounding:

• API Gateway + Lambda + Cognito — authenticated API layer
• Bedrock Knowledge Base — RAG over S3 documents (PDFs, policies, manuals)
• OpenSearch Serverless — vector store for 1M+ document chunks
• Guardrails — PII redaction + topic blocking for compliance
• DynamoDB — conversation history for multi-turn context
• CloudWatch — latency, token usage, error metrics

An agent that can perform actions on behalf of the user:

• Customer support agent — look up orders, process refunds, update accounts
• Research agent — search documents, summarize findings, create reports
• DevOps agent — check service health, read CloudWatch logs, trigger remediation

Use Case	Best Service	Why
Generative AI (text, chat, code)	Bedrock	Serverless, multiple models, built-in RAG
Custom ML model (train from scratch)	SageMaker	Full training control, custom algorithms
Image recognition (pre-trained)	Rekognition	No ML needed, simple API
Text analysis (sentiment, entities)	Comprehend	Pre-built NLP, no model selection
Self-hosted open models	SageMaker endpoints	Custom containers, dedicated GPUs

Mistake	Why It's Bad	Fix
Using Opus for everything	5-10x more expensive than needed	Route by complexity — Haiku for simple, Sonnet for complex
Fine-tuning before trying RAG	Expensive, slow, RAG works better	Start with Knowledge Bases first
No guardrails in production	Unfiltered content to users	Always enable Guardrails — PII + content filters
Ignoring token costs	Long system prompts = waste	Use Prompt Caching, keep prompts concise
Not using VPC endpoints	Data traverses public internet	Enable VPC endpoints for Bedrock API calls