What is Amazon Bedrock?

Most companies that want to use AI models already run on a cloud. Signing a new contract with each model maker, sending data outside the cloud account and managing a new set of API keys is the slow part. Amazon Bedrock removes it on AWS: the models are called from inside your AWS account, like any other AWS service.

One service, many models

Bedrock serves more than 100 models from many makers: Anthropic's Claude, OpenAI's GPT models, Amazon's Nova, DeepSeek, Kimi from Moonshot AI, MiniMax, xAI's Grok and others. You call them with the API you already know: Anthropic's Messages API for Claude, OpenAI's Responses or Chat Completions API for GPT, or Bedrock's own Converse API, which takes the same request shape for every model.

import boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")
response = client.converse(
    modelId="anthropic.claude-sonnet-5-5",
    messages=[{"role": "user", "content": [{"text": "Summarise this ticket: ..."}]}],
)

What the AWS account adds

• Permissions. Who may call which model is an IAM policy, like access to any other AWS resource.
• Logs. Calls appear in CloudTrail and CloudWatch, in your account.
• Privacy from the model maker. Models run in accounts the Bedrock team operates. AWS says the model makers have no access to them, so they cannot see your prompts or answers.
• Networking. Traffic can stay on a private connection inside AWS.
• Billing. Usage goes on the AWS bill, which matters to companies with AWS spending commitments.

What it costs you

Features arrive on the maker's own API first, and some never reach Bedrock: for Claude, the server-side tools, the Files API and the Batches API are among those missing. Limits are per model and per Region, counted in requests and tokens per minute, and on Bedrock the max_tokens you ask for is held against the limitRate LimitThe most an API lets you send in a period, counted in requests and tokens per minute. Past it, requests fail with a 429 error until capacity comes back, so code waits and retries. until the request ends. Retirement dates are Bedrock's own and can differ from the maker's; see Model DeprecationModel DeprecationThe notice that a model will be switched off on a set date. Until then it still answers; after its retirement date, every request to it fails, so code must move to a newer model first..

Beyond calling a model

Bedrock is also a family of services around the models: knowledge bases for retrieval, guardrails, model evaluation, fine-tuning and distillation, batch inference, and AgentCore for running agentsAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. in production with their memory, tools and permissions.

Our articles compare Claude on Bedrock with the Claude API and explain Amazon Bedrock AgentCore.