Amazon Bedrock: the complete 2026 guide.

IWhat Amazon Bedrock is, and why AWS built it

Amazon Bedrock is a managed AWS service that gives you foundation models from many providers through a single API, with no infrastructure to run. You send a prompt, get a response, and pay for the tokens.

Before Bedrock, a team that wanted a large language model in production had two hard options. Rent GPUs, download open weights and run the serving stack yourself, or send your data to a single model vendor's API. Bedrock removes both problems. AWS runs the inference fleet behind the API. And your prompts and outputs are not used to train the base models, are not shared with the model providers, and stay in your AWS account and the Region you call.

The second reason is choice without lock-in. No single model is best at everything, and through the Converse API, switching models is often a one-line change to a model ID.

The third reason is that Bedrock is an application platform, not just a model gateway. AWS adds managed building blocks that most teams would otherwise build themselves: Agents, Knowledge Bases for RAG, Guardrails, fine-tuning and distillation, Flows, Prompt Management, model evaluation, and AgentCore for running agents in production. Section V covers each.

IIThe model catalog

Breadth is Bedrock's defining feature. On 10 October 2026 the <a href="https://aws.amazon.com/bedrock/pricing/">Bedrock pricing page</a> listed models from 19 providers: AI21 Labs, Amazon, Anthropic, Cohere, DeepSeek, Google, Luma AI, Meta, MiniMax AI, Mistral AI, Moonshot AI, NVIDIA, OpenAI, Qwen, Stability AI, TwelveLabs, Writer, xAI and Z AI. Versions and Regions change often, so confirm the live list in the Bedrock console.

A few patterns help before reading the table. Anthropic's Claude family is the most widely used set of models on Bedrock for reasoning, coding and agent work, in tiers that trade cost against capability. Amazon's Nova family covers low-cost text and multimodal models (Micro, Lite, Pro, Premier), the newer Nova 2 generation (Nova 2 Lite, with Nova 2 Pro and Omni in preview), images (Canvas), video (Reel) and speech (Sonic). Open-weight models from Meta, Mistral, DeepSeek, Qwen and others can be compared side by side without self-hosting. Embedding models from Amazon and Cohere power retrieval.

You do not have to pick one provider for a whole application. A common pattern sends cheap, high-volume calls (classification, extraction, routing) to a small model such as Nova Micro, Nova Lite or Claude Haiku, and escalates only the hard reasoning steps to a frontier model. All of it runs through the same Converse API and lands on one bill.

IIIHow access works: models, IAM, Regions and the Converse API

Getting to a first call is short. Access to many models is on by default, IAM controls who can call what, you pick a Region, and you call models with the Converse API.

IVHow Bedrock pricing works

Bedrock has no platform fee and no minimum. You pay for what you run, through a few levers. The table shows list prices for Amazon's own models, checked on the AWS pricing page on 10 October 2026.

On-demand is the default: a price per million input tokens and per million output tokens, set per model, with no commitment. Output costs several times more than input. Image and video models bill per image or per second of video. AWS also offers service tiers (Standard, Flex, Priority and Reserved) that trade price against latency and capacity.

Batch processes large sets of prompts asynchronously at a 50% lower price than on-demand for supported models. For offline work (overnight document processing, bulk classification, embedding a corpus), batch should be the default.

Provisioned Throughput reserves model capacity by the hour, with discounts for 1-month and 6-month commitments. It gives steady throughput and is needed to serve some custom models. It only pays off at high, steady volume.

Prompt caching stores repeated context (a long system prompt, a document, tool definitions) so you do not pay full input price to process it again on every call. For Nova 2 models, AWS prices cache reads at 75% less than on-demand input. Expect extra charges for model customization and for services around the model, such as the vector store behind a Knowledge Base.

VThe feature suite: what turns model access into a product

Teams standardize on Bedrock rather than a bare model API because of the managed layer around the models. Each feature below would otherwise be a project of its own.

• Agents: Bedrock Agents let a model plan multi-step tasks and call your APIs and Lambda functions (look up an order, file a ticket, run a query), with orchestration and tool calls managed for you. See Amazon Bedrock Agents.
• AgentCore: Amazon Bedrock AgentCore is AWS's platform for running agents in production with any framework and any model. It covers runtime, a gateway to tools and MCP servers, identity and access control for tool calls, observability and evaluations (AgentCore page, checked 10 October 2026).
• Knowledge Bases: Managed RAG: point it at documents in S3, and Bedrock chunks them, creates embeddings, stores them in a vector database, and grounds answers in the retrieved passages, with citations. See Bedrock Knowledge Bases and RAG on AWS.
• Guardrails: A safety layer that filters harmful content, blocks denied topics, redacts or blocks PII, and adds grounding checks, applied the same way across any model. See Bedrock Guardrails.
• Fine-tuning and distillation: Customize a base model on your own labeled data, or distill a large model's behavior into a smaller, cheaper one. See Bedrock fine-tuning.
• Flows: A visual builder for chaining prompts, models, Knowledge Bases, Agents and logic into one deployable workflow. See Bedrock Flows.
• Prompt Management and model evaluation: Version prompts as resources, and compare candidate models on your own data before you commit.

These pieces compose. A typical production assistant uses a Knowledge Base for grounding, wraps the model in Guardrails, takes actions through an Agent or AgentCore, keeps its prompts in Prompt Management, and was chosen with model evaluation, all under one IAM and billing boundary.

VISecurity, privacy, compliance and data residency

For most companies the deciding factor in favor of Bedrock is governance, not a benchmark. Bedrock is designed so that adopting generative AI does not mean loosening your data controls.

Your data is yours. Prompts, responses and anything you send to Bedrock are not used to train the base models and are not shared with the model providers. Content is encrypted in transit and at rest, you can use your own KMS keys, and a request is processed in the Region you call. Cross-region inference stays inside a defined geography, such as EU Regions for an EU profile.

It plugs into your existing controls. Reach Bedrock privately from your VPC through PrivateLink, govern access with IAM, record every call in CloudTrail, and log full payloads for audit. Bedrock is in AWS's main compliance programs (such as SOC, ISO 27001, HIPAA eligibility and PCI DSS, depending on Region); confirm scope for your Region in AWS Artifact. This is the most common reason teams pick Bedrock over sending sensitive data to a public model endpoint.

VIIBedrock vs SageMaker vs calling model APIs directly

Bedrock is not the only way to run AI on AWS, and it is not always the right one. The rule comes down to how much of the ML stack you want to own and how much data governance matters.

Bedrock answers "I want existing foundation models through a managed, secure API with the least operations." Amazon SageMaker answers "I need to own the ML lifecycle: my own model or architecture, custom training, control of serving, or classical ML." A vendor's API directly answers "I want this vendor's newest model the moment it ships, and my data can leave AWS."

They work together: Bedrock for the generative AI layer, SageMaker for a custom forecasting or vision model, in one account. Details on Bedrock vs SageMaker and Bedrock vs OpenAI.

VIIIGetting started, and who pays for the build

A first Bedrock application takes a day, not a quarter. The harder problem is the bill once the application is real and traffic grows.

The fast path: pick one workhorse model and one embeddings model in your Region, attach an IAM policy scoped to those model ARNs, make a first Converse call, then point a Knowledge Base at a folder of documents in S3 if you need grounding, and wrap it in a Guardrail. You now have a grounded, governed assistant without a single GPU.

Then the cost arrives: generative AI is cheap per call and expensive in aggregate. The levers from section IV keep it in check: route easy calls to small models, run offline work as batch, turn on prompt caching, and reserve capacity only once volume is steady. See Bedrock pricing, prompt caching and batch inference.