What Is Amazon Bedrock? Models, Pricing, and Alternatives

Amazon Bedrock started as a way to call a handful of foundation models through one AWS API. In 2026 it's the single door to more than 100 models from 18 providers, and the list changes almost monthly. That scale is the appeal and the catch: Bedrock gives you the models and the infrastructure, but you still have to build the agent that uses them.

This guide covers what Bedrock actually is today, which models it supports in 2026, what it costs, what you can build with it, and where a dedicated agent platform like Voiceflow fits.

What Is Amazon Bedrock?

Amazon Bedrock is a fully managed AWS service for building generative AI applications on top of foundation models (FMs). Instead of hosting a model yourself, you call it through one API, and AWS handles the serving, scaling, and security.

You can use models as-is, ground them in your own data with retrieval-augmented generation, or fine-tune them on your examples. Bedrock also ships higher-level building blocks: Agents for multi-step tasks, Knowledge Bases for managed RAG, Guardrails for safety filters, and Flows for chaining steps together. Because it's serverless and lives inside AWS, it connects to the services you already run (S3, Lambda, IAM) without separate infrastructure to manage.

The short version: Bedrock is the model-access and infrastructure layer. What you build on top of it is up to you.

What Models Does Amazon Bedrock Support in 2026?

This is where Bedrock has changed the most. The original lineup was six providers; today the catalog spans roughly 18 providers and well over 100 individually addressable models, all behind the same API. The headline additions:

• Amazon Nova. AWS's own family, launched at the end of 2024 and now central to Bedrock: Nova Micro, Lite, Pro, and Premier for text, Nova Canvas and Reel for images and video, and Nova Sonic, a speech-to-speech model that went generally available in early 2026. The Nova tiers are among the cheapest options on the platform.
• Anthropic Claude. Still the most popular family on Bedrock, now through the Claude 4 generation including Opus-class reasoning models. See our Anthropic guide for the full lineup.
• OpenAI. GPT-5.5 and GPT-5.4, plus Codex and OpenAI's open-weight gpt-oss models, are now available on Bedrock, which would have been unthinkable a year ago.
• DeepSeek. DeepSeek-R1, the open reasoning model, is available and priced low enough to matter for high-volume work.
• Meta Llama, Mistral, Cohere, AI21, Stability AI, Google Gemma, Qwen, Writer, Luma AI, TwelveLabs, and NVIDIA Nemotron round out the rest, covering text, code, image, and video.

The practical benefit isn't any single model. It's that you can switch between them by changing a model ID rather than re-architecting, which is the same argument for keeping your options open across large language models generally. If you're weighing specific families, our Mistral, GPT-5, and DeepSeek vs ChatGPT breakdowns go deeper.

How Much Does Amazon Bedrock Cost?

Bedrock is usage-based, and there's no flat fee. The 2026 pricing comes in a few models you'll mix depending on workload:

• On-demand. You pay per token, now quoted per million tokens, split into input and output. This is the default for most apps.
• Batch. Submit large jobs asynchronously for roughly half the on-demand price. Good for offline processing where latency doesn't matter.
• Provisioned throughput. Reserve dedicated capacity by the hour for steady, high-volume traffic or custom models.
• Customization. Fine-tuning and continued pre-training carry their own training and storage costs.

Prompt caching and Guardrails are billed separately, and prices vary widely by model. To anchor the range: Amazon Nova Lite runs around $0.06 per million input tokens, Nova Pro around $0.80 input and $3.20 output, while frontier Claude and OpenAI models cost considerably more. The lesson for budgeting is that your model choice, not the platform, dominates the bill. If token-based pricing is new to you, our primer on tokens explains what you're paying for. Always check the AWS Bedrock pricing page for current per-model rates before you commit.

What You Can Build With Amazon Bedrock

Bedrock is general-purpose, so the range is wide:

• Customer support agents that answer questions and take action. DoorDash built a generative self-service support experience on Claude models in Bedrock to handle interactions across customers, merchants, and Dashers.
• Internal copilots grounded in company docs through a Knowledge Base, so answers come from your content rather than the model's training data.
• Event and product assistants. The PGA Tour built an assistant that surfaces real-time schedules, stats, and player information for fans.
• Content and media generation using Nova, Stability, and Luma for text, images, and video from the same platform.

Most of these lean on Bedrock Agents and Knowledge Bases to move past single-turn responses toward AI agents that reason over multiple steps.

Building an Agent on Bedrock: What It Does and Doesn't Do

Here's the honest boundary. Bedrock gives you model access, RAG plumbing, and the AWS infrastructure to run it. It does not give you the conversation design, the deployment surface across chat and voice, or the monitoring you need once real users show up. That's the work of building a production agentic AI application, and on raw Bedrock it lands on your engineering team: wiring Lambda functions, managing state, standing up a front end, and building your own evaluation and observability.

For a team with AWS depth and time, that's a reasonable trade. For most teams that want an agent in production without assembling the stack themselves, it's a lot of undifferentiated work.

Amazon Bedrock Alternatives

Bedrock isn't the only way to reach foundation models or build agents. The main options:

• Amazon SageMaker is AWS's own machine-learning platform, aimed at data scientists who train and deploy custom models. It's lower-level than Bedrock and overkill if you only want to call an FM.
• Microsoft Azure AI and Google Vertex AI are the equivalent managed model services on the other two clouds. If you're already committed to one, start there. See our Azure and Vertex AI guides.
• IBM watsonx leans toward regulated enterprises that want governance and data controls built in, per our IBM Watson breakdown.
• Voiceflow is the alternative when what you actually want is the agent, not the model plumbing.

Amazon Bedrock vs Voiceflow

These two solve different problems, and the honest framing is that they're layers, not competitors. Bedrock is where the models live. Voiceflow is where you build, deploy, and watch the agent that uses them.

• Model-agnostic by design. Voiceflow runs across OpenAI, Anthropic, and Google, and you can point it at Bedrock-hosted models by API. You switch as cost and quality change instead of committing to one provider's stack. Our roundup of the best AI chatbots goes wider if you're shortlisting.
• Real agent primitives. Workflows handle the deterministic steps that must run the same way every time; Playbooks handle open-ended reasoning. You design both on a visual canvas rather than in application code.
• Grounded and multi-channel. A built-in Knowledge Base keeps answers on your content, and the same agent ships to website chat, messaging, and voice, with a clean human handoff when it reaches its limit.
• Built for production. Observability gives you turn-by-turn transcripts, Evaluations score conversations against your own criteria, and Environments let you test in staging first. If you've evaluated standalone tooling like Arize or Braintrust, the difference is that this is wired into the same place you build. For sensitive data, Voiceflow runs with SOC 2 Type 2 and PII masking.

Teams like Turo, StubHub International, Sanlam Studios, and Trilogy build their agents this way. A common setup is to keep Bedrock as the model layer and build the agent in Voiceflow on top of it, so you get AWS's model access and a platform designed for shipping agents.

{{blue-cta}}

Frequently Asked Questions

Is Amazon Bedrock free?

No. Bedrock is usage-based with no flat fee, so you pay for what you use, mostly per token. There's no free tier, but the cheapest models (like Amazon Nova Lite at roughly $0.06 per million input tokens) make small-scale testing inexpensive. Costs scale with your model choice and traffic.

What models does Amazon Bedrock support?

As of 2026, Bedrock offers 100+ foundation models from about 18 providers through one API, including Amazon Nova, Anthropic Claude, OpenAI GPT-5.x and gpt-oss, Meta Llama, DeepSeek, Mistral, Cohere, AI21, Stability AI, Google Gemma, Qwen, Writer, Luma AI, TwelveLabs, and NVIDIA Nemotron. They span text, code, image, and video.

What is the difference between Amazon Bedrock and SageMaker?

Bedrock is a managed service for calling and customizing existing foundation models through an API, aimed at developers building generative AI apps. SageMaker is AWS's broader machine-learning platform for training and deploying your own custom models, aimed at data scientists. If you want to use an FM, choose Bedrock; if you want to build models from scratch, choose SageMaker.

Is Amazon Bedrock an LLM?

No. Bedrock is not a model itself. It's the AWS service that gives you access to many large language models (and image and video models) from different providers. The LLMs, like Claude or Nova, run on Bedrock.

What is the best Amazon Bedrock alternative?

It depends on what you need. For model access on another cloud, Azure AI or Google Vertex AI are the direct equivalents. For actually building and deploying a customer-facing agent, Voiceflow is the most common alternative, since it's model-agnostic, supports chat and voice, and includes the observability and evaluation tooling Bedrock leaves to you.

{{blue-cta}}