Amazon Bedrock

AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is now Bedrock Agents Classic, in maintenance for new customers after July 30, 2026. Net-new agent builds should use Bedrock AgentCore. Full matrix: lifecycle roundup.

Definition

Amazon Bedrock is a fully managed generative AI service that exposes foundation models (FMs) through a unified API inside your AWS account. As of August 2026, Bedrock offers roughly 100 serverless models from Amazon, Anthropic, Meta, Mistral, Cohere, OpenAI, xAI, and others — including Claude Sonnet 5 (June 30, 2026), Claude Opus 5 (July 24, 2026), GPT-5.6 Sol, Terra, and Luna (July 2026), Grok 4.3, Fable 5 (June 9, 2026), Opus 4.8 (May 28, 2026), Sonnet 4.6, the Amazon Nova family (Micro through Premier, plus Canvas and Reel), and open-weight options such as Llama. Bedrock also provides Knowledge Bases (managed RAG with S3 Vectors or OpenSearch backends), Agents Classic (maintenance for new customers after July 30, 2026), Guardrails, Prompt Flows, Data Automation, and the bedrock-mantle endpoint (OpenAI Responses/Chat Completions and Anthropic Messages APIs). Net-new multi-step agents should use Bedrock AgentCore. Model inference and customer content stay within your AWS environment under the standard AWS shared responsibility model.

When to use it

• Production GenAI without GPU fleet management — pay per token instead of operating inference clusters.
• Model diversity in one integration — route classification to Nova Micro, high-volume chat to GPT-5.6 Luna, agentic coding to Claude Sonnet 5, and maximum reasoning to Claude Opus 5 from the same codebase (Converse API or bedrock-mantle).
• Grounded Q&A over private documents via Knowledge Bases and the Converse / RetrieveAndGenerate APIs.
• Regulated workloads requiring Guardrails, CloudTrail logging, VPC endpoints, and IAM-scoped model access.
• Multi-step agents that call APIs, query knowledge bases, and retain session memory — extend with Bedrock AgentCore for managed runtime infrastructure.

When not to use it

• Lowest cost at very high sustained throughput on a single fixed model — self-hosted inference on SageMaker or EC2 can win economically after careful modeling.
• Fully custom architecture not available on Bedrock — exotic fine-tuning, non-supported frameworks, or on-device inference need different stacks.
• Single-turn tasks solvable without an LLM — traditional search, rules engines, or classical ML may be cheaper and more deterministic.

Tips

• Enable model access per region for each provider in the Bedrock console before deployment — missing access shows as obscure API errors in CI.
• Apply Guardrails in production on both input and output regardless of model vendor.
• Run model evaluation on your own prompt set before standardizing on Opus-class models — Sonnet 5, Sonnet 4.6, or Nova Pro often meet the quality bar at lower cost.
• Use cross-region inference profiles where available for resilience when a single region throttles.
• Separate dev and prod IAM policies and Knowledge Base data sources to prevent test prompts from retrieving production documents.

Gotchas

Serious

• No Guardrails on customer-facing endpoints — models can leak training-adjacent patterns, generate harmful content, or echo PII from prompts.
• Throttling during launch — new workloads need quota headroom and exponential backoff; cross-region profiles help but do not eliminate limits.
• Treating all models as interchangeable — tool use, context length, multimodal input, and fine-tuning support vary by model ID.

Regular

• Hardcoding model IDs — AWS releases new versions; pin in config and plan upgrade testing when IDs deprecate.
• Skipping citation review in RAG — Knowledge Bases reduce hallucination but do not eliminate wrong-chunk retrieval; validate sources in UI.
• Logging full prompts with secrets — CloudWatch logs may capture API payloads; redact credentials and PHI at the application layer.

Official references

• What is Amazon Bedrock?
• Supported foundation models
• Amazon Bedrock Knowledge Bases
• Guardrails for Amazon Bedrock

Related FactualMinds content

• Generative AI on AWS Bedrock
• AWS Bedrock Consulting
• Fine-Tuning vs RAG on AWS Bedrock: When to Use Each
• RAG Pipeline