What Amazon Bedrock is
Amazon Bedrock is AWS's managed platform for accessing, customizing, and deploying AI models. Think of it as a single front door into a marketplace of AI: instead of building and running your own AI infrastructure, you call Bedrock's API and it handles the servers, scaling, and security behind the scenes. The models themselves come from multiple providers — Anthropic, OpenAI, Mistral, Hugging Face, and others — all accessible through the same AWS account and billing relationship you already have.
Why it matters
For most businesses, the hard part of using AI isn't finding a good model — it's connecting that model to real systems, keeping data secure, meeting compliance requirements, and not rebuilding everything every time a better model comes out. Bedrock is designed to solve exactly those problems. It sits inside AWS's existing security and compliance envelope, which means enterprises can use frontier AI without routing sensitive data through unfamiliar third-party infrastructure.
The scale of adoption reflects this: tens of thousands of enterprises run Claude on Bedrock, including household names like Pfizer, Intuit, Lyft, Salesforce, and Snowflake. Amazon itself uses Bedrock to power Alexa+, its upgraded voice assistant. Lyft reported an 87% reduction in customer service resolution time after deploying Claude via Bedrock.
What you can do with it
Beyond basic model access, Bedrock offers several enterprise-grade capabilities:
- Fine-tuning: You can train a model on your own data to specialize it for your use case. Claude 3 Haiku, for example, can be fine-tuned on Bedrock — one early adopter improved a content moderation task from 81.5% to 99.6% accuracy while cutting token usage by 85%.
- Agents and stateful workflows: Bedrock supports multi-step AI agents that can remember context, use tools, and carry out long tasks. OpenAI's stateful runtime environment for agents — which manages memory, tool connections, and permissions across a session — runs on Bedrock.
- Guardrails and Knowledge Bases: Built-in safety filters and the ability to connect models to your own documents and databases.
- GovCloud support: Claude models on Bedrock are approved for FedRAMP High and DoD Impact Level 4 and 5 workloads in AWS GovCloud regions, making them available to federal agencies and defense organizations handling sensitive government data.
The model lineup
Bedrock's catalog spans the spectrum from small, fast, cheap models to frontier-class ones:
- Anthropic Claude: The full Claude family, from Haiku (fast and affordable at $1/$5 per million tokens) through Sonnet to Opus 4.8 (the current flagship for complex coding and agentic tasks). Claude Code — Anthropic's autonomous coding tool — also defaults to Claude Opus 4.8 on Bedrock.
- OpenAI: Models available via a Responses API endpoint, plus the stateful agent runtime.
- Mistral: Including Mistral Large 3 and the Ministral edge model series, all open-weight under Apache 2.0.
- Hugging Face: Open-weight and proprietary models through the Bedrock model marketplace.
The Anthropic relationship
Anthropic and Amazon have the deepest partnership in the Bedrock ecosystem. Amazon has invested a total of over $8 billion in Anthropic, with a 10-year agreement committing over $100 billion to AWS technologies and securing up to 5 gigawatts of compute capacity on Amazon's custom Trainium chips for training and running Claude. Anthropic engineers work directly on the low-level software that makes those chips run efficiently. The full Claude Platform — including every tier of Claude model — is available within AWS as a result of this deal.
Where it's heading
Bedrock is evolving from a model-access layer into a broader AI infrastructure platform. The addition of OpenAI's stateful agent runtime signals that AWS wants Bedrock to be where enterprise AI agents live — not just where models are called. The GovCloud approvals open a large and previously hard-to-reach market. And as more providers join the catalog, Bedrock's value as a single managed interface only grows.
