Overview
Amazon Bedrock is a fully managed generative AI foundation platform launched by AWS, with the core positioning of "one API to call multiple foundation models". It currently integrates Anthropic Claude Opus 4.8 / Sonnet 5, Meta Llama, Mistral, Cohere, Stability AI, AWS's self-developed Amazon Nova series models (including Nova Pro/Lite/Micro/Canvas/Reel), and — newly added — OpenAI GPT-5.6 models (Sol, Terra, and Luna), serving as the official entry point for enterprises to build AI applications using the AWS ecosystem.
Compared to the OpenAI API, Bedrock's differentiation lies in multi-model + enterprise-level governance: Bedrock Agents (build Agent workflows), Knowledge Bases (managed RAG), Guardrails (content security + PII filtering), Bedrock Studio (low-code development), Model Evaluation (model evaluation), and Fine-tuning are all natively integrated. Capabilities such as data staying within the AWS account, VPC private network, IAM fine-grained permissions, and SOC/HIPAA/FedRAMP compliance are the core reasons for its widespread adoption among Fortune 500 companies.
The platform has since extended well beyond model access into a full agent runtime layer. Amazon Bedrock AgentCore now anchors production agent workloads: teams migrate multi-model agents (including LangGraph customer support agents and multi-model healthcare agents) from self-managed ECS/Fargate onto AgentCore Runtime and Gateway (2026/9/3, 2026/9/18), optimize agent system prompts by turning production traces into suggested configuration changes validated before rollout (2026/9/16), and govern agent tool access through a four-scope maturity model — connect, control, catalog and harden (2026/8/21). Web Search on AgentCore added runtime domain and publish-date filtering, giving developers per-call control over which web sources agents consult and the publication timeframe of retrieved content (2026/8/19). AgentCore also underpins applied builds such as Natera's automated voice agent for mobile phlebotomy appointment scheduling, orchestrated with Amazon Transcribe and Amazon Polly (2026/8/26), and context-aware personal assistants that accumulate context across conversations (2026/10/6).
On the knowledge side, the Amazon Bedrock Managed Knowledge Base supports enterprise agentic retrieval in which an agent reasons, routes across multiple knowledge bases, and returns results (2026/8/31), while access control for RAG over sources like SharePoint, Google Drive, and Confluence is addressed together with Amazon Quick (2026/10/7). Amazon Quick itself is being made enterprise-ready through automated, auditable promotion of resources — agents, action connectors, knowledge bases, flows, and spaces — from development to production AWS accounts (2026/10/5). Cost governance at scale is also covered: Jamf built real-time, per-user spend enforcement for Amazon Bedrock using IAM Customer Managed Policies (2026/9/1).
Key Features
- Multiple Foundation Models: Unified access to Claude Opus 4.8 / Sonnet 5, Llama, Mistral, Cohere, Stability, Amazon Nova, and OpenAI GPT-5.6 models (Sol, Terra, Luna), etc.
- Bedrock Agents: Natively build multi-step Agent workflows, automatically orchestrate function calls and external APIs
- Amazon Bedrock AgentCore: Production agent runtime with Runtime and Gateway for migrating agentic workloads, including LangGraph and multi-model agents, off self-managed infrastructure (2026/9/3, 2026/9/18)
- AgentCore System Prompt Optimization: Turns production traces into suggested configuration changes, then validates them before rollout (2026/9/16)
- AgentCore Gateway Tool Governance: Governed, auditable agent access to enterprise tools, organized around a four-scope maturity model: connect, control, catalog and harden (2026/8/21)
- Web Search on AgentCore: Runtime domain and published-date filtering, with per-request filters giving per-call control over web sources and publication timeframe (2026/8/19)
- Knowledge Bases (Managed RAG): Connect S3 / Aurora / OpenSearch for automatic vector retrieval, out-of-the-box RAG; the Managed Knowledge Base supports agentic retrieval that reasons and routes across multiple knowledge bases (2026/8/31)
- Guardrails Security Barriers: Content filtering, PII masking, topic blocking, hallucination detection, essential for enterprise compliance
- Model Evaluation: Supports automatic and human evaluation, enabling horizontal comparison of different models on business tasks
- Private Network + IAM: Supports VPC endpoints and PrivateLink, data stays within the AWS account, IAM fine-grained permissions
- Fine-tuning / Continued Pre-training: Fine-tune or continue pre-training models with enterprise private data, with weights still owned by the customer
Use Cases
- AI application foundation for Fortune 500 / finance / healthcare and other highly regulated industries
- Teams needing a unified API to call multiple models, avoiding single-vendor lock-in
- Enterprises already heavily using the AWS ecosystem (S3/Lambda/SageMaker)
- Production-grade scenarios requiring end-to-end capabilities of Agent + RAG + Guardrails
- Projects with PII, HIPAA, FedRAMP, data sovereignty requirements
- Teams migrating agentic workloads from self-managed infrastructure to a managed agent runtime (2026/9/3, 2026/9/18)
- Enterprise RAG over permissioned knowledge sources such as SharePoint, Google Drive, and Confluence (2026/10/7)
- Organizations needing real-time, per-user spend enforcement for generative AI (2026/9/1)
- Voice-first agents, including appointment scheduling built with Amazon Transcribe and Amazon Polly (2026/8/26)
Pros
- Unified access to multiple models, extremely low switching cost
- Most comprehensive enterprise governance capabilities (Guardrails / IAM / VPC / Audit) in the industry
- Amazon Nova series offers cost-effective proprietary models
- Deep integration with S3 / Lambda / SageMaker / Aurora
- Comprehensive compliance certifications (SOC 2, HIPAA, FedRAMP, PCI DSS)
- Model weights and data remain entirely within the customer's AWS account
- AgentCore provides a production agent runtime with Runtime, Gateway, prompt optimization, and governed tool access
- Managed Knowledge Base supports agentic retrieval across multiple knowledge bases
- Supports real-time, per-user spend enforcement via IAM Customer Managed Policies
Pricing
Billed per model per token, no fixed subscription. Reference prices: Claude Opus 4.8 input ~$15/M, output ~$75/M; Claude Sonnet 5 input $2-3/M, output $10-15/M; Llama series $0.20-$3/M; Amazon Nova Micro lowest input $0.035/M, output $0.14/M. Additional charges include Knowledge Bases retrieval fees, Guardrails invocation fees, and Fine-tuning training fees. AWS Free Tier offers some free credits. For spend control at scale, Bedrock supports real-time, per-user enforcement built on IAM Customer Managed Policies, as implemented by Jamf (2026/9/1).
Summary
Amazon Bedrock is the "official enterprise-grade entry point" for AI applications—unified multi-model API + the most complete governance capabilities + AWS ecosystem, making it the top choice for Fortune 500 companies to deploy generative AI. Its model lineup now spans Anthropic Claude, Meta Llama, Mistral, Cohere, Stability AI, Amazon Nova, and OpenAI GPT-5.6 (Sol, Terra, Luna), and its agent story has matured into the Amazon Bedrock AgentCore runtime with Runtime, Gateway, prompt optimization, governed tool access, and Web Search filtering. If you are building AI applications on AWS, Bedrock is almost the default choice; independent developers and startup teams can start with OpenAI API / OpenRouter and switch to Bedrock when compliance needs arise.
Version History
- Anchor Browser and AgentCore: agents that autonomously pay for x402-protected resources (2026-10-13): The AWS Generative AI Innovation Center Partner Agent Factory and Anchor Browser show how to build AI agents on Amazon Bedrock AgentCore that autonomously pay for x402-protected web resources and APIs, with no human step in the payment path.
- 以防万一你错过了：2026年9月为AI构建者带来了什么 (2026-10-09): 2026年9月Amazon Bedrock、Amazon Bedrock AgentCore和Strands最新更新的月度回顾：更广泛的模型选择、更快的无服务器代理并内置评估，以及自动化
- 重新思考使用Amazon Quick和Amaz进行RAG的访问控制 (2026-10-07): Enterprise RAG unlocks insights from knowledge sources like SharePoint, Google Drive, and Confluence, but those sources carry complex permissions. Learn how Amazon Quick and Amazon Bedrock Knowledge B
- Building a context-aware AI assistant on AgentCore and OpenC (2026-10-06): Off-the-shelf AI assistants forget you between conversations. This post shows how to build a personal assistant that accumulates context using OpenClaw on Amazon Bedrock AgentCore runtime, with AgentC
- Making Amazon Quick enterprise-ready: Automated, auditable c (2026-10-05): 将Amazon Quick资源（代理、操作连接器、知识库、流程和空间）从开发AWS账户推广到生产AWS账户一直是一项手动且容易出错的工作。这篇文章展示了如何
- 将接近Astra的智能带入日常工作，借助GPT-6.1 (2026-09-29): GPT-6.1 Sol 现已在 Amazon Bedrock 上正式可用，为频繁运行的编码、计算机使用和专业工作负载带来更强的推理能力。
- Migrate multi-model AI agents to Amazon Bedrock AgentCore (2026-09-18): Migrate a multi-model healthcare AI agent from self-managed Amazon ECS with AWS Fargate to Amazon Bedrock AgentCore runtime, preserving triple-model orchestration and vector-enhanced knowledge retriev
- Optimize agent system prompts with Amazon Bedrock AgentCore (2026-09-16): AgentCore optimization turns production traces into suggested configuration changes, then validates them before rollout. This technical companion to the launch post explains system prompt optimization
- Migrate agentic workloads to Amazon Bedrock AgentCore (2026-09-03): An agent running in a notebook is not an agent running in production. This article walks through migrating a LangGraph customer support agent to Amazon Bedrock AgentCore in two phases: first to Runtime, then to Gateway,
- Accessing OpenAI models on Amazon Bedrock from Australia wit (2026-09-02): Australian teams can now access OpenAI GPT-5.6 Sol, Terra, and Luna models on Amazon Bedrock with global cross-Region inference from the Asia Pacific (Sydney) and Asia Pacific (Melbourne) Regions. Thi
- Tokenomics at scale: How Jamf built real-time spend enforcem (2026-09-01): As generative AI adoption scales, cost governance becomes a top challenge. Learn how Jamf built real-time, per-user spend enforcement for Amazon Bedrock using IAM Customer Managed Policies, an Amazon
- Build observable enterprise agentic retrieval using Managed (2026-08-31): This post builds an enterprise agentic retrieval solution on the Amazon Bedrock Managed Knowledge Base and Amazon Bedrock AgentCore. An agent reasons, routes across multiple knowledge bases, and retur
- Introducing OpenAI models on Amazon Bedrock for in-country i (2026-08-27): Amazon Bedrock now supports the OpenAI GPT-5.6 models, Terra and Luna, in India with India geographic cross-Region inference. If you have local data processing requirements, you can now use these mode
- Natera’s intelligent appointment scheduling with Amazon Bedr (2026-08-26): Learn how Natera built an automated voice agent on Amazon Bedrock AgentCore that lets patients book mobile phlebotomy appointments through natural conversation. The post covers the dual-WebSocket bridge architecture, the orchestration of Amazon Bedrock AgentCore with Amazon Transcribe and Amazon Polly, and the design patterns used to manage long-running tasks. It also discusses the challenges encountered, such as handling WebSocket reconnections and managing concurrent sessions, along with the solutions implemented to ensure a seamless patient experience.
- Democratizing institutional knowledge: Building an AI-powere (2026-08-24): Learn how to build a customizable, smart-caching knowledge management system on AWS that captures and delivers institutional (tribal) knowledge through a voice-first AI avatar. The accelerator uses Am
- Govern AI agent tool access with Amazon Bedrock AgentCore Ga (2026-08-21): Give your AI agents governed, auditable access to enterprise tools without stitching together infrastructure. This article lays out a four-scope maturity model (connect, control, catalog and harden) to help get there.
- Introducing cross-Region inference for OpenAI GPT-5.6 models (2026-08-20): Amazon Bedrock now offers OpenAI GPT-5.6 models (Sol, Terra, and Luna) in more than 25 AWS Regions with cross-Region inference. Learn how US geographic and global inference profiles route requests for
- Domain and publish date filters for Web Search on AgentCore (2026-08-19): Web Search on Amazon Bedrock AgentCore now supports runtime domain and published-date filtering. New per-request filters give developers per-call control over which web sources their agents consult and the publication timeframe of the content they retrieve.