What is Amazon Bedrock (AWS Bedrock)? Models, pricing, agents and what breaks in production
If you already know Bedrock and need the agent layer, start with AgentCore architecture, components and pricing. This page is the overview.
By Jerzy Kopaczewski, Devopsity (AWS consulting partner). Service and pricing facts were checked against AWS documentation on 21 November 2026.
What Amazon Bedrock is
Amazon Bedrock is AWS’s managed service for using foundation models. Instead of provisioning GPUs, downloading weights and running an inference server, you send a request to an API and AWS runs the model. AWS documents it as secure, enterprise-grade access to high-performing foundation models from leading AI companies, for building and scaling generative AI applications.
Is Bedrock part of AWS? Yes. “AWS Bedrock” and “Amazon Bedrock” are the same product, one service inside the AWS cloud alongside EC2, S3 or Lambda. Access is controlled with IAM, traffic can stay on AWS networking, and usage appears on your normal AWS bill. That is the main reason teams already on AWS choose it over calling a model vendor directly: security review, logging and cost allocation work the way they already do.
How Bedrock works
You work with Bedrock through the bedrock-runtime endpoint. AWS’s own documentation recommends it for new applications and shows several ways to call a model: the Converse API (one consistent request shape across models), the Invoke API, and compatibility APIs such as Messages, Responses and Chat Completions for some providers’ models. The practical consequence is that switching a model is often a change of model ID, not a rewrite.
Around inference, Bedrock adds the features most teams end up needing: retrieval over your own data, safety controls, evaluation and customisation. These are covered below. A short checklist of what to settle before you build:
- Model access and Region. Models are enabled per account and offered per Region. Check both before you design around a specific model.
- Cross-Region inference. AWS offers Global and Geo cross-Region inference to raise throughput and reduce per-token cost for supported models, with the data-residency trade-off that implies.
- Quotas. Requests and tokens per minute are limited per account and per model, which is where most production surprises start.
The models on Bedrock
AWS’s documentation describes more than 100 foundation models from providers including Amazon (Nova), Anthropic (Claude), DeepSeek, Moonshot AI, MiniMax, OpenAI and xAI. The pricing page also lists providers such as AI21 Labs, Cohere, Google, Meta, Mistral AI, NVIDIA, Qwen, Stability AI and TwelveLabs. The catalogue changes quickly, and new models arrive on Bedrock regularly, so treat any fixed list, including this one, as a snapshot and check the current model list.
Choosing a model is also choosing a price and a capacity profile: the same task can differ several-fold in cost between providers and model sizes. For a worked comparison of what that means at scale, see SageMaker vs Bedrock vs self-hosted EKS: LLM cost.
What you can build with it
| Capability | What it is for |
|---|---|
| Model inference | Calling a foundation model for chat, summarisation, extraction, code or image tasks, on demand or in batch. |
| Knowledge Bases | Retrieval-augmented generation over your own documents. The vector store behind it has its own cost; see RAG on AWS: patterns and costs. |
| Guardrails | Content and safety controls applied to prompts and responses. |
| Model customisation | Fine-tuning, continued pre-training and distillation, per AWS's documentation, when a prompt is not enough. |
| Evaluation, prompt routing and optimisation | Comparing models, routing requests and tuning prompts; each is a separately priced feature. |
| Agents and AgentCore | Building agents that call tools and keep memory. AgentCore is the runtime layer: architecture and pricing, and how it compares to self-hosted and framework options. |
AWS keeps adding tools on top, such as a built-in Web Search tool that grounds OpenAI models on Bedrock in current web information. If you are comparing agent platforms across clouds, see AgentCore vs Vertex vs Foundry. Related products often confused with Bedrock are covered in What is Amazon Bedrock Studio and SageMaker Unified Studio vs Bedrock.
Planning a Bedrock workload and want a second opinion on cost and architecture?
Book a free 30-min call
How AWS Bedrock pricing works
There is no single “Bedrock price”. AWS states that pricing depends on the modality, the provider and the model, and the bill is the sum of several metered pieces.
- Tokens, by model. Text models bill for input and output tokens at a rate set per model. Output is usually the more expensive side, so verbose answers and long reasoning cost more than short ones.
- Service tiers. AWS offers Standard, Flex, Priority and Reserved tiers. Standard is the default pay-as-you-go path; Reserved suits steady, high-volume traffic that cannot tolerate capacity errors; the others trade price against latency and priority. Pick the tier per workload, not per account.
- Batch inference. For select models, AWS prices batch inference at 50% below on-demand. Anything that does not need an immediate answer, such as nightly document processing, belongs here.
- Features are billed separately. Knowledge Bases, Guardrails, Model Evaluation, Data Automation, Intelligent Prompt Routing, Prompt Optimization and Web Search each have their own pricing, so a “simple” RAG assistant bills for the model, the retrieval service and the vector store.
- Agent loops multiply tokens. An agent that plans, calls tools and re-reads context uses far more tokens than a single call. Cost the loop, not the prompt. Agent runtime pricing is in AgentCore pricing.
Bedrock, SageMaker or self-hosted?
- Bedrock fits when you want managed access to existing models with the least operational work, and your data and workloads are already on AWS.
- SageMaker fits when you need to train, fine-tune heavily or host your own models. See SageMaker Unified Studio vs Bedrock.
- Self-hosting on EKS fits at high, steady volume where you want control of cost and placement, and you can staff the operations. The numbers are in our LLM inference cost comparison.
What breaks in production
Most Bedrock content stops at “what is it”. These are the errors teams actually hit, each with a fix in the Devopsity runbook library:
| Symptom | Usual cause | Runbook |
|---|---|---|
| 503 ServiceUnavailableException | Temporary capacity strain on a shared model | Bedrock 503: unable to process your request |
| 429 ThrottlingException | Account quota (requests or tokens per minute) | InvokeModel 429 throttling |
| "Too many tokens per day" | Daily token quota exhausted | Tokens-per-day limit |
| AccessDenied on a model | Legacy or deprecated model, or access not enabled | Model AccessDenied |
| Knowledge Base sync fails | S3 permissions on the data source | Knowledge Base S3 access denied |
| AgentCore runtime timeouts | Invocation exceeds runtime limits | Runtime invocation timeout |
When it is worth bringing in an external partner
As an AWS consulting partner, Devopsity sees these as the Bedrock situations where an outside team earns its fee:
- Choosing the layer. Bedrock, AgentCore, SageMaker or self-hosting solve different problems; the wrong pick is an expensive detour.
- Cost modelling before rollout. Token, retrieval and agent-loop costs are easy to underestimate. A realistic model avoids the first-month surprise, and Devopsity’s FinOps work covers ongoing control.
- Resilience. Retries, backoff, cross-Region inference and reserved capacity decide whether a throttled or overloaded model becomes an outage.
- Security and data residency. IAM scoping, Guardrails and Region choice need to be settled before sensitive data reaches a model.
Summary
Amazon Bedrock (AWS Bedrock) is AWS’s managed way to use foundation models from many providers, with the surrounding features for retrieval, safety, customisation and agents.
Three things Devopsity would have you remember:
- It is one AWS service, not a separate vendor. Security, networking, logging and billing work the way they do for the rest of your AWS estate, which is its main advantage over calling a model vendor directly.
- The bill is shaped, not fixed. Model, tier, batch versus on-demand, separately priced features and agent loops all move the cost, so model your own usage first.
- Plan for capacity and quota errors. 503s and 429s are normal at scale; retries, quotas and reserved capacity are part of the design, not an afterthought.
Frequently asked questions about AWS Bedrock
What is AWS Bedrock?
Is Bedrock part of AWS? What is the difference between AWS and AWS Bedrock?
What is AWS Bedrock AI used for?
Which models are available on Amazon Bedrock?
How much does Amazon Bedrock cost?
Why am I getting a 503 or 429 error from Bedrock?
Should I use Bedrock, SageMaker or self-host?
Building on Amazon Bedrock?
Book a free 30-minute call with Devopsity. We help teams choose between Bedrock, SageMaker and self-hosting, model the real token and retrieval costs before rollout, and harden Bedrock workloads against throttling and capacity errors.