Amazon Bedrock Guide: Models, Agents, Guardrails and Pricing

What Is Amazon Bedrock?

Not long ago, adding generative AI to a business application meant hiring ML engineers, provisioning GPU clusters, negotiating with several model providers and building inference pipelines from scratch. Today, however, it can start with one API call, and Amazon Bedrock is the AWS service behind that call.

Amazon Bedrock is a fully managed service from Amazon Web Services that gives you secure, serverless access to foundation models (FMs) from the world’s leading AI companies: all through a single unified API. So rather than managing infrastructure, you pick a model, send a prompt, and get a response. Whether you need text generation, image creation, code assistance, or autonomous AI agents, Bedrock handles the compute, scaling, and security behind the scenes.

Since general availability in September 2023, Bedrock has grown from a model marketplace into a broad enterprise AI platform, with agents, safety controls, managed retrieval and cost tools on top of the models themselves.

What separates Bedrock from calling model providers directly is the enterprise layer: Guardrails for safety, Knowledge Bases for managed RAG, AgentCore for agents, cost tools such as Intelligent Prompt Routing and Model Distillation, and compliance coverage that includes SOC, PCI DSS, HIPAA eligibility, FedRAMP High and ISO 27001. So teams can build AI applications that are secure, compliant and cost-aware from the start.

Amazon Bedrock by the Numbers

Adoption has followed. Bedrock is now a default choice for many AWS customers building generative AI, from startups with a first chatbot to large enterprises running high-volume inference.

Where Amazon Bedrock Fits in the AWS AI Ecosystem

AWS offers multiple AI and machine learning services, and understanding where Bedrock sits relative to them is essential for making the right architectural choice:

• Amazon Bedrock: Fully managed access to pre-trained foundation models via API. It is designed for teams building generative AI applications without managing infrastructure. Choose Bedrock when you want to use existing models, not train your own from scratch.
• Amazon SageMaker: A complete ML platform for building, training, and deploying custom models. Or choose SageMaker when you need to train proprietary models on your own datasets or require full control over the ML lifecycle.
• Amazon Q: An AI assistant for business users and developers. Built on top of Bedrock’s infrastructure, Amazon Q provides a ready-to-use conversational interface rather than requiring you to build one.

In practice, many organizations use Bedrock and SageMaker together: Bedrock for inference and application integration, SageMaker for training and custom model work. The two complement each other.

How Amazon Bedrock Works

Understanding Bedrock’s architecture also helps you make better decisions about model selection, cost management, and application design. At its core, Bedrock is an abstraction layer: it sits between your application and the foundation models, handling inference, scaling, and security so you do not have to.

Traditionally, integrating an AI model into an application required provisioning GPU-backed instances, managing model weights, implementing autoscaling, building request queuing, and handling model version management. So teams spent months on infrastructure before writing a single line of application logic. Bedrock eliminates this entire layer. Instead, you interact with models through API calls, and AWS handles everything behind the endpoint, from GPU allocation to load balancing to failover.

Foundation Models and the Unified API

Bedrock offers a large catalog of foundation models from many providers, including Anthropic (Claude), OpenAI open-weight models, Meta (Llama), Mistral AI, Amazon (Nova and Titan), Cohere, AI21 Labs, Stability AI, DeepSeek, MiniMax, Moonshot and Qwen, across text, code, image, video and audio.

Also, all models are accessible through a single, unified API. Whether you are calling Claude for complex reasoning or Nova Micro for lightweight classification, the integration pattern remains the same. Bedrock offers three API interfaces:

• Converse API: The recommended interface for conversational workloads. It also provides a consistent request/response format across all text models, regardless of provider. Handles model-specific formatting differences automatically.
• InvokeModel API: The lower-level interface that passes model-specific payloads. Offers maximum flexibility but requires provider-specific formatting for each model family.
• OpenAI-compatible endpoints: Bedrock also offers endpoints compatible with OpenAI’s API format, so teams already built on that format can move with fewer code changes.

Choosing the Right Model for Your Workload

With so many models available, selecting the right one can feel overwhelming. However, the decision framework is straightforward when you map task complexity to model ability. You should consider these categories:

• Simple classification, extraction, and routing tasks: Use lightweight models like Amazon Nova Micro or Nova Lite. As a result, these handle routine tasks at a fraction of the cost of frontier models: often delivering comparable quality for structured, well-defined prompts.
• General-purpose reasoning, summarization, and content generation: Mid-tier models like Claude Sonnet, Llama, or Mistral offer an excellent balance of ability and cost. So most production workloads settle in this tier.
• Complex reasoning, multi-step analysis and agentic workflows: Frontier models such as Claude Opus give the highest quality on hard tasks. Reserve them for work where the quality justifies the price.
• Image and video generation: Amazon Nova Canvas (images), Nova Reel (video), and Stability AI models serve creative and media workloads with per-asset pricing.
• Embeddings and search: Cohere Embed and Amazon Titan Embeddings power semantic search, recommendation engines, and RAG pipelines. So choose based on language coverage and dimensionality needs.

Inference Modes and Service Tiers

Bedrock offers multiple service tiers that let you balance cost, latency, and throughput based on your workload requirements. The five tiers are:

• Standard (On-Demand): Pay-per-token with no commitment. It is best for variable workloads, experimentation, and development environments.
• Priority: Higher throughput with a time-based commitment. It is also designed for production workloads that need guaranteed availability.
• Flex: Lower-cost access for workloads that can tolerate slightly higher latency. Delivers the same savings as batch processing but through the regular API: no workflow restructuring needed.
• Reserved (Provisioned Throughput): Dedicated capacity with a term commitment. As a result, it provides predictable performance for high-volume, latency-sensitive production systems.
• Batch: Asynchronous processing for non-time-sensitive workloads. Submit prompts as a file, receive results within 24 hours at roughly half the on-demand cost.

Architecture Under the Hood

When your application sends a request to Bedrock, several things happen behind the scenes. First, the request is authenticated via IAM and routed to the appropriate model endpoint. Next, if you have configured Guardrails, the input is evaluated against your safety policies before reaching the model. Then, the model processes the request and generates a response. Finally, if Guardrails are active, the output is also evaluated before being returned to your application.

Bedrock is fully serverless, with no instances to provision, no GPUs to manage and no capacity planning. AWS states that your prompts and outputs are not used to train the base models or shared with model providers, and data is encrypted in transit and at rest. As a result, Bedrock fits regulated industries where privacy matters.

Core Amazon Bedrock Features and Abilities

Since its launch, Bedrock has evolved far beyond a simple model API. As of 2026, it is a comprehensive AI platform with abilities spanning agent orchestration, safety controls, knowledge management, model customization, and cost optimization. Below are the features that matter most for production deployments.

Amazon Bedrock AgentCore: Building Production AI Agents

AgentCore is Bedrock’s platform for building, deploying and operating AI agents at scale. Unlike a single prompt and response, agents reason about a task, break it into steps, call tools, query databases and act across your systems. Note that the original Amazon Bedrock Agents service, now called Bedrock Agents Classic, moved to maintenance mode and closed to new customers on July 30, 2026, so AgentCore is the path for new agent projects.

AgentCore covers what enterprise agents need in production: runtime hosting, memory, identity, a gateway for tools including MCP (Model Context Protocol) servers, observability and policy controls that check an agent’s actions outside its reasoning loop before they reach tools or data.

It works with popular agent frameworks, and AWS has published customer examples across marketing, financial services and software.

In other words, AgentCore is framework-agnostic: you can build agents using any orchestration framework (LangChain, AutoGen, AWS Strands) and deploy them on AgentCore for production-grade management. This decoupling means your agent logic is portable, while AgentCore handles the operational complexity of scaling, monitoring, and securing agents in production.

Guardrails for Responsible AI

Guardrails is Bedrock’s safety layer for checking user inputs and model outputs against your policies. According to AWS, Guardrails can block up to 88% of harmful content, and Automated Reasoning checks can verify correct model responses with up to 99% accuracy.

For example, Guardrails lets you define denied topics (subjects the model must not discuss), content filters (hate speech, violence, sexual content, insults), word blocklists, PII redaction rules, and grounding checks that reduce hallucinations. Also, you can apply Guardrails across all models in your account consistently, so safety policies follow the organization, not individual applications.

Guardrails also works independently of the model you choose. Whether you switch from Claude to Llama or from Nova to Mistral, your safety policies stay in force, which matters for compliance in financial services and healthcare. AWS has also cut Guardrails prices by up to 85% for content filters and denied topics, so applying safety checks to every interaction is more affordable.

Knowledge Bases and RAG Pipelines

Knowledge Bases enable Retrieval Augmented Generation (RAG): the technique of grounding model responses in your organization’s actual data. Rather than relying solely on a model’s training data, Knowledge Bases connect your S3 data sources, automatically chunk and embed the documents, store the vectors in a managed vector store, and retrieve relevant context at query time.

As a result, the model’s responses are grounded in your proprietary data, product catalogs, internal documentation, compliance policies, customer records, rather than general knowledge. So this dramatically reduces hallucinations and makes the output genuinely useful for enterprise applications.

Also, Knowledge Bases support multiple vector store backends including Amazon OpenSearch Serverless, Amazon Aurora, Pinecone, and Redis Enterprise. Also, Bedrock Data Automation handles document parsing for complex file types, PDFs, images, spreadsheets, and multi-page documents, extracting structured data that the Knowledge Base can index and retrieve. For organizations building internal search tools, customer support bots, or compliance assistants, Knowledge Bases provide the critical bridge between general AI ability and organization-specific knowledge.

Model Customization and Fine-Tuning

While Bedrock’s pre-trained models handle most use cases well out of the box, some applications require domain-specific customization. Bedrock supports two customization approaches. First, fine-tuning lets you train a private copy of a model on your labeled dataset to improve performance on specific tasks. Second, continued pre-training lets you expose the model to your unlabeled domain data to improve its general understanding of your business context.

For Amazon’s own models, Nova Forge lets organizations build customized versions of Nova with their own data.

Intelligent Prompt Routing and Cost Controls

One of Bedrock’s most impactful cost optimization features is Intelligent Prompt Routing. Rather than sending every request to your most expensive model, this feature automatically routes prompts to the most cost-effective model within a family based on complexity. Simple queries go to the lighter model; complex ones go to the more capable one.

According to AWS, Intelligent Prompt Routing can reduce costs by up to 30% without hurting quality. Together with Model Distillation (distilled models up to 500% faster and 75% cheaper), prompt caching (up to 90% lower cost on cached tokens) and batch processing (roughly 50% cheaper), Bedrock gives teams a full set of tools for controlling AI spend.

Talk to an ExpertChoosing models and guardrails on Bedrock? Design your generative AI stack with us.

Pricing Model and Cost Optimization

Bedrock uses a token-based, pay-per-use pricing model for text workloads. A token is approximately four characters or 0.75 words in English. Input tokens (your prompt) are always cheaper than output tokens (the model’s response), typically by a factor of 3-5x. Similarly, for image and video models, pricing is per image generated or per second of video.

Understanding the Five Pricing Modes

Rather than listing specific dollar amounts that change over time, here is how each pricing mode works and when to use it:

• On-Demand: Pay per token with no commitment. Most flexible but highest per-token cost. It is best for experimentation, development, and variable workloads.
• Batch: Submit prompts asynchronously and receive results within 24 hours. As a result, it is roughly 50% cheaper than on-demand for the same model. So it is ideal for content generation, data processing, and any non-real-time workload.
• Flex: Same savings as batch but through the standard API: no workflow restructuring required. Suitable for workloads that can tolerate slightly higher latency without going fully asynchronous.
• Priority: Higher throughput guarantee with a time-based commitment. It is also designed for production workloads that need consistent performance and availability.
• Provisioned Throughput: Dedicated capacity purchased in model units with a term commitment. Provides the most predictable latency and highest throughput, but at a fixed cost regardless of usage. So reserve this for high-volume, latency-critical production systems only.

Also, Bedrock supports cross-region inference that automatically routes traffic across AWS Regions to avoid capacity constraints: with no additional charge beyond the source region’s pricing. So this provides a valuable availability guarantee for production workloads without adding cost complexity.

Strategies to Reduce Your Amazon Bedrock Bill

Based on our experience deploying Bedrock for enterprise clients, these strategies deliver the most significant cost reductions:

• Right-size your model selection: Prices vary enormously between the cheapest and most expensive models on Bedrock, so using a frontier model for simple classification or extraction wastes budget. Match the model to the task; lightweight models handle routine work well at a fraction of the cost.
• Enable Intelligent Prompt Routing: Automatically routes each request to the most cost-effective model in a family. As a result, it saves up to 30% on inference costs with minimal quality impact.
• Use prompt caching: For applications that share common context across requests (system prompts, reference documents), prompt caching avoids reprocessing the same tokens repeatedly. Reductions of up to 90% on cached portions are possible.
• Distill models for production: If you have a high-volume, well-defined task, Model Distillation creates a smaller, faster model trained on the outputs of a larger one: up to 500% faster and 75% cheaper.
• Leverage batch and Flex tiers: Any workload that does not require sub-second responses should use batch (50% savings) or Flex for cost efficiency.

Amazon Bedrock Security and Compliance

For enterprise adoption, security is not optional: it is the prerequisite. Amazon Bedrock was architecturally designed for regulated industries from the ground up, and its security posture reflects that commitment.

Data Protection Architecture

Bedrock provides several security guarantees. First, AWS states that your prompts and completions are not used to train the base models and are not shared with model providers. Second, data is encrypted in transit (TLS) and at rest (AWS KMS). Third, AWS PrivateLink gives private connectivity between your VPC and Bedrock, so inference traffic stays off the public internet. Fourth, your prompts, responses and custom model data stay isolated from other customers.

Bedrock also supports identity-based access control through IAM policies, so you control which users, roles and applications can call specific models. For chargeback, tag workloads and track inference costs by team, project and application in Cost Explorer.

Also, comprehensive monitoring and logging abilities support governance and audit requirements. CloudWatch integration provides metrics on model invocations, token usage, latency, and error rates. CloudTrail records all API calls for security auditing. Together, these controls create a complete audit trail from request to response: a requirement for many regulated industries.

Compliance Certifications

Amazon Bedrock is in scope for compliance programs including SOC 1, SOC 2, SOC 3, PCI DSS, ISO 27001, FedRAMP High and CSA STAR Level 2, it is HIPAA eligible and it supports GDPR requirements. AWS also offers uncapped intellectual property (IP) indemnity for copyright claims arising from the output of covered Amazon AI models.

What’s New in Amazon Bedrock (2025-2026)

Bedrock has changed quickly. These are the main updates of recent releases:

Real-World Amazon Bedrock Use Cases

Amazon Bedrock applies to almost every industry. Financial services, healthcare, retail, media and technology companies use it to solve real business problems beyond proofs of concept. These are the use cases we implement most often for enterprise clients:

Amazon Bedrock vs Azure OpenAI Service

If you are evaluating enterprise AI platforms across cloud providers, here is how Amazon Bedrock compares with Microsoft’s Azure OpenAI Service:

Making the Right Platform Decision

The choice depends on your cloud ecosystem and AI strategy. If you are an AWS-native organization wanting access to the broadest model selection with maximum flexibility, Bedrock is the stronger choice. Conversely, if your stack runs on Azure and you primarily need GPT-family models with deep Microsoft integration, Azure OpenAI is the natural fit.

Bedrock’s main advantage is model choice without dependence on one provider. If a provider changes pricing, has an outage or deprecates a model, Bedrock users can often switch to an alternative with a configuration change. Bedrock’s cost toolkit (distillation, routing, caching, batch and Flex) is also broad, which makes costs easier to control at scale.

Getting Started with Amazon Bedrock

Surprisingly, you can make your first Bedrock API call in under five minutes. Here is a step-by-step walkthrough.

Enabling Model Access

First, navigate to the Amazon Bedrock console in the AWS Management Console. Click Model access in the left navigation. Next, select the models you want to enable: by default, no models are accessible until you explicitly grant access. Also, some models (like Claude) require you to accept the provider’s end-user license agreement. Once enabled, there is no cost for enabling access: you only pay when you make inference calls.

Your First API Call

Below is a minimal Python example using the Converse API with Boto3. Before running this code, ensure you have the AWS CLI configured with appropriate credentials and the boto3 library installed (pip install boto3):

import boto3

# Initialize the Bedrock Runtime client
client = boto3.client('bedrock-runtime', region_name='us-east-1')

# Send a request using the Converse API
response = client.converse(
    modelId='anthropic.claude-sonnet-4-6-v1',
    messages=[
        {
            'role': 'user',
            'content': [{'text': 'Explain Amazon Bedrock in two sentences.'}]
        }
    ]
)

# Print the response
print(response['output']['message']['content'][0]['text'])

Or if your team already uses the OpenAI Python SDK, you can use Bedrock’s OpenAI-compatible endpoint with only small changes to your client setup:

from openai import OpenAI

client = OpenAI()  # Uses Bedrock's Mantle endpoint

response = client.chat.completions.create(
    model="anthropic.claude-sonnet-4-6-v1",
    messages=[
        {"role": "user", "content": "Explain Amazon Bedrock in two sentences."}
    ]
)

print(response.choices[0].message.content)

Or for a no-code experience, the Bedrock console includes playgrounds for text, chat, and image generation where you can test models interactively before writing any code.

Best Practices and Common Pitfalls

Based on our experience building production AI applications on Bedrock for enterprise clients across industries, these are the practices that consistently determine success or failure.

Recommendations for Production Deployment

• Start with the Converse API: Unless you have a specific reason to use InvokeModel, the Converse API provides a consistent interface across all text models and simplifies future model switching. Also, it handles provider-specific formatting differences automatically, so your code remains clean as you experiment with different models.
• Next, enable Guardrails from day one: Retrofitting safety controls after launch is significantly harder than building them in from the start. So define your content policies, PII redaction rules, and denied topics before deploying to production.
• Implement cost monitoring early: Use CloudWatch metrics, the new IAM cost allocation tags, and S3 logging to track token consumption by team, project, and application. So you can identify cost spikes before they become budget problems.

Architectural Best Practices

• Design for model portability: Structure your application so that switching models requires a configuration change, not a code rewrite. This is one of Bedrock’s greatest architectural advantages: do not negate it by hardcoding model-specific logic throughout your application.
• Test with the playground before coding: Use the Bedrock console playgrounds to evaluate model quality, test prompts, and compare outputs before committing to integration work. This saves significant development time by validating assumptions early.
• Build evaluation pipelines: Use Bedrock’s model evaluation abilities to systematically compare models on your actual tasks. Rather than selecting a model based on benchmarks alone, evaluate it on representative samples of your real-world data to ensure it meets your quality requirements at the right cost point.

Our SolutionEnterprise AI Adoption and AI Readiness

Frequently Asked Questions About Amazon Bedrock

References

1. Models, APIs, Knowledge Bases, Guardrails and Data Privacy: AWS — Amazon Bedrock User Guide
2. Model Distillation, Prompt Caching and Intelligent Prompt Routing Savings: AWS — Amazon Bedrock Pricing
3. Guardrails 88% Harmful Content, Automated Reasoning 99%: AWS — Amazon Bedrock Guardrails
4. Bedrock Agents Classic Moving to Maintenance July 30, 2026: AWS What’s New — AWS Service Availability Updates, June 2026
5. AgentCore: AWS — Amazon Bedrock AgentCore