Amazon Bedrock Pricing 2026: Tiers, Models and Real Costs

Last Updated: November 8, 2026

Summary: Amazon Bedrock pricing is pay as you go: you pay per million input and output tokens for the model you call, with no monthly fee, and rates differ by model and region. AWS then adjusts that base price by service tier: Flex runs at a 50% discount, Priority at a 75% premium, and batch jobs on supported models cost 50% less than on-demand. Add-ons such as Knowledge Bases and Guardrails are billed separately, and Claude models carry a 10% premium on regional endpoints over global ones.

Bedrock is Amazon Web Services' managed service for running foundation models from Anthropic, Meta, Mistral, DeepSeek, OpenAI, Amazon and others through one API. Searches for its pricing have climbed sharply this year as more companies move AI workloads onto their existing AWS accounts, but the official price page is long, split by provider, and easy to misread.

This guide pulls the numbers that matter for a business budget into one place: the tier multipliers, per-token rates for popular models, what Claude costs on Bedrock versus buying it from Anthropic, the add-on fees, and a worked monthly cost example. All prices were checked against AWS and Anthropic documentation on November 8, 2026.

Amazon Bedrock charges for what you use, mainly the tokens a model reads and writes, with no subscription or seat fee. Each model has its own input and output rate, usually quoted per million tokens, and the rate can vary by AWS region.

On top of the base token price, Bedrock pricing depends on four choices:

One billing detail surprises finance teams: third-party models such as Claude are sold through AWS Marketplace. The Amazon Bedrock model card for Claude Sonnet 5 says its charges appear on the AWS bill and in AWS Cost Explorer under the model provider, not under Amazon Bedrock. Anyone building a cost report should filter for both.

Bedrock's service tiers change the price of the same model by a fixed percentage. According to the official Amazon Bedrock pricing page, Flex is priced at a 50% discount to Standard, Priority at a 75% premium, and batch inference on select models at 50% below on-demand.

Not every model supports every tier. The AWS Claude Sonnet 5 model card lists Standard as supported and Priority, Flex and Reserved as not supported for that model, so the 50% Flex discount cannot be assumed for Claude. Check the model card for each model before you plan savings around a tier.

Provisioned Throughput is where Bedrock bills can jump. AWS's own pricing example shows one Cohere model unit at $39.60 an hour, which works out to $29,462.40 for a 31-day month. That makes sense only for workloads big enough to keep the capacity busy.

Popular non-Claude models on Bedrock cost from about $0.15 to $2.00 per million tokens, depending on the model and whether the tokens are input or output. These are Standard on-demand list prices from the Amazon Bedrock pricing page, checked November 8, 2026.

Output tokens usually cost more than input tokens, so tasks that write long answers cost more than tasks that read long documents and reply briefly. For a broader view of how these rates compare with direct APIs, see our DeepSeek pricing guide and our monthly AI Pricing Index.

Claude's price on Bedrock is set by AWS, not Anthropic, so the Bedrock price page is the final word. Anthropic's pricing documentation says partner platforms, including Bedrock, "have independent regional pricing" and sends readers to AWS for official Bedrock rates.

Three rules from official sources shape what Claude costs on Bedrock:

For reference, here are Anthropic's own list prices for its current models when bought directly through the Claude API:

There is now a second way to buy Claude inside AWS. Anthropic's pricing documentation describes Claude Platform on AWS, which bills through AWS Marketplace in Claude Consumption Units at $0.01 each. Usage is rated at the same per-model rates as the Claude API, metered hourly and invoiced monthly. A company that wants Anthropic's list prices and features on its AWS bill can use that route instead of Bedrock.

Feature gaps matter too. Anthropic's Bedrock guide lists features not supported on Bedrock, including the Message Batches API, server-side tools such as web search and code execution, the Files API, the MCP connector, Agent Skills and automatic prompt caching. Our Claude API pricing guide covers the first-party options in more detail.

Sameer's take: Claude is my favorite of all the AI tools, and I use it every day.

Bedrock's add-on features are billed on top of model tokens, and for retrieval-heavy apps they can cost more than the model itself. These are the main meters listed on the Amazon Bedrock pricing page.

AWS's worked examples show how quickly the add-ons scale. An agentic research assistant making 100,000 calls a month over 50 GB of data costs $850 a month in Knowledge Bases fees alone: $250 for storage, $400 for agentic retrieval and $200 for the underlying Retrieve calls. A customer support chatbot screening 3,000 text units an hour with content filters and denied topics costs $0.90 an hour in Guardrails fees, according to AWS. By our estimate, extending that hourly figure to round-the-clock use over a 30-day month comes to roughly $650.

Bedrock Flows are cheap by comparison: AWS's example of a flow run 960 times a month totals $0.84. AgentCore, AWS's agent platform, is not priced on the Bedrock pricing page, but AWS notes that standard Gateway tool invocation charges apply if you call a managed Knowledge Base through AgentCore Gateway, and standard CloudWatch rates apply if observability is enabled.

A support assistant handling 10,000 conversations a month costs from about $5 to $110 a month in model fees at current list prices, depending on the model. The example assumes each conversation sends 3,000 input tokens (instructions, history and a help-center excerpt) and returns 500 output tokens, for 30 million input and 5 million output tokens a month.

The math is simple: 30 times the input rate plus 5 times the output rate. Mistral Large 3, for example, is 30 x $0.50 plus 5 x $1.50, or $22.50. Two adjustments apply in real life. First, if you run Claude on a Bedrock regional endpoint for data residency, add 10% to Bedrock's global Claude price. Second, prompt caching can cut the input side sharply when every conversation reuses the same long instructions, since Claude cache hits cost a fraction of the base input rate.

The model fee is often not the biggest line. If this assistant also searched a Knowledge Base on every conversation, 10,000 standard retrievals would add $10, plus $5 per GB stored each month.

Pick Bedrock if you want many model providers, AWS-native guardrails and retrieval, and one AWS bill. Pick Claude Platform on AWS if Claude is your main model and you want Anthropic's list prices while still paying through AWS.

If you are comparing clouds, our Azure OpenAI pricing guide covers Microsoft's side, and our Gemini API pricing and OpenAI API pricing guides cover the direct APIs.

Is Amazon Bedrock free to use?
No. Amazon Bedrock has no monthly fee, but every model call is billed per token, and features such as Knowledge Bases and Guardrails have their own charges. You pay only for what you use.

How much cheaper is Bedrock batch inference?
Batch inference on select Bedrock models is priced 50% lower than on-demand inference, according to AWS. Not every model supports batch, so check the model's page first.

Does Claude cost the same on Bedrock as on the Claude API?
Not necessarily. Anthropic says Bedrock has independent regional pricing and points to AWS's Bedrock pricing page for official rates, and Claude on a Bedrock regional endpoint costs 10% more than on a global endpoint. Claude Platform on AWS bills at the same rates as the Claude API.

What are the Amazon Bedrock service tiers?
Bedrock offers Standard, Flex, Priority and Reserved tiers. Flex costs 50% less than Standard, Priority costs 75% more, and Reserved pricing is set with your AWS account team.

Where do Claude charges appear on an AWS bill?
Claude models are third-party models sold through AWS Marketplace. AWS's Claude Sonnet 5 model card says the charges appear on your AWS bill and in Cost Explorer under the model provider, not under Amazon Bedrock.

Amazon Bedrock pricing comes down to three levers: the model you pick, the tier you run it on and the add-ons you switch on. Models such as Ministral 8B and Mistral Large 3 cost cents to a few dollars per million tokens, Flex and batch halve the bill where supported, and Priority adds 75%.

For Claude, check AWS's own rate card, budget 10% extra for regional endpoints, and compare against Claude Platform on AWS, which charges Anthropic's list prices through the same AWS bill. Then price the add-ons, because Knowledge Bases and Guardrails can outgrow the model fee on a busy app.

Written by Sameer Khan with AI assistance, fact-checked against the sources linked above.