Amazon Bedrock Pricing

Pricing overview

Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models (FMs) through a single API, along with a broad set of capabilities you need to build generative AI applications with security, privacy, and responsible AI.

With Amazon Bedrock, you will be charged for model inference and customization. You have a choice of two pricing plans for inference: 1. On-Demand and Batch: This mode allows you to use FMs on a pay-as-you-go basis without having to make any time-based term commitments. 2. Provisioned Throughput: This mode allows you to provision sufficient throughput to meet your application's performance requirements in exchange for a time-based term commitment.

Pricing Models

With the On-Demand mode, you only pay for what you use, with no time-based term commitments. For text-generation models, you are charged for every input token processed and every output token generated. For embeddings models, you are charged for every input token processed. A token comprises a few characters and refers to the basic unit of text that a model learns to understand the user input and prompt. For image-generation models, you are charged for every image generated.

Cross-region inference: On-Demand mode also supports cross-region inference for some models. It enables developers to seamlessly manage traffic bursts by utilizing compute across different AWS Regions and get higher throughput limits and enhanced resilience. There's no additional charge for using cross-region inference and the price is calculated basis the region you made the request in (source region).

With Batch mode, you can provide a set of prompts as a single input file and receive responses as a single output file, allowing you to get simultaneous large-scale predictions. The responses are processed and stored in your Amazon S3 bucket so you can access them at a later time. Amazon Bedrock offers select foundation models (FMs) from leading AI providers like Anthropic, Meta, Mistral AI, and Amazon for batch inference at a 50% lower price compared to on-demand inference pricing. Please refer to model list here.

Latency-optimized inference for foundation models in Amazon Bedrock delivers faster response times for models and helps improve responsiveness for your generative AI applications. You can use the use latency-optimized inference for Anthropic's Claude 3.5 Haiku model, and Meta's Llama 3.1 405B and 70B models. As verified by Anthropic, with latency-optimized inference on Amazon Bedrock, Claude 3.5 Haiku runs faster on AWS than anywhere else. Additionally, with latency-optimized inference in Bedrock, Llama 3.1 405B and 70B runs faster on AWS than any other major cloud provider. Learn more here.

With the Provisioned Throughput mode, you can purchase model units for a specific base or custom model. The Provisioned Throughput mode is primarily designed for large consistent inference workloads that need guaranteed throughput. Custom models can only be accessed using Provisioned Throughput. A model unit provides a certain throughput, which is measured by the maximum number of input or output tokens processed per minute. With the Provisioned Throughput pricing, you are charged by the hour, you have the flexibility to choose between 1-month or 6-month commitment terms.

Custom Model Import allows you to leverage your prior model customization investments within Amazon Bedrock and consume them in the same fully-managed manner as Bedrock’s existing hosted foundation models. You can import custom weights for supported model architectures and serve the custom model using On-Demand mode. There is no charge to import a custom model to Bedrock. Once you import a model, you will be able to access it on-demand without requiring to perform any control plane action. You are only charged for model inference, based on the number of copies of your custom model required to service your inference volume and the duration each model copy is active, billed in 5-minute windows. A model copy is a single instance of an imported model ready to serve inference requests. The price per model copy per minute depends on factors such as architecture, context length, AWS Region, compute unit version (hardware generation), and is tiered by model copy size.

With Amazon Bedrock, you can customize FMs with your data to deliver tailored responses for specific tasks and your business context. You can fine-tune models with labeled data or using continued pretraining with unlabeled data. For customization of a text-generation model, you are charged for the model training based on the total number of tokens processed by the model (number of tokens in the training data corpus x the number of epochs) and for model storage charged per month per model. An epoch refers to one full pass through your training dataset during fine-tuning or continued pretraining. Inferences using customized models are charged under the Provisioned Throughput plan and requires you purchase Provisioned Throughput. One model unit is made available with no commitment term for inference on a customized model. You will be charged for the number of hours you use in the first model unit for custom model inference. If you want to increase your throughput beyond one model unit, then you must purchase a 1-month or 6-month commitment term.

Tools

Amazon Bedrock Guardrails helps you to implement customized safeguards and responsible AI policies for your generative AI applications. It provides additional customizable safety protections on top of the native protections offered by FMs. It is the only responsible AI capability offered by a major cloud provider that helps enable customers to build and customize safety, privacy, and truthfulness protections for their generative AI applications in a single solution, and it works with all FMs in Amazon Bedrock, as well as fine-tuned models. Bedrock Guardrails can also be integrated with Amazon Bedrock Agents and Amazon Bedrock Knowledge Bases to build generative AI applications aligned with your responsible AI policies. Additionally, it offers an ApplyGuardrail API to help evaluate user inputs and model responses generated by any custom or third-party FM outside of Bedrock.

Amazon Bedrock Knowledge Bases provides a fully managed end-to-end retrieval-augmented generation (RAG) workflow, enabling FMs and agents to access contextual information from your company's private data sources. This allows them to deliver more relevant, accurate, and customized responses. You can securely connect FMs and agents to multiple data sources such as Amazon S3, Confluence, Salesforce, and SharePoint. If you don't have an existing vector database, Amazon Bedrock creates an Amazon OpenSearch Serverless vector store for you. Alternatively, you can specify an existing vector store in supported databases like Amazon OpenSearch Serverless, Pinecone, Redis Enterprise Cloud, Amazon Aurora, and MongoDB. You can also fine-tune retrieval and ingestion to achieve better accuracy across use-cases with advanced parsing options for unstructured data, data chunking options like custom chunking, or built-in chunking strategies including default, fixed size, no chunking, hierarchical chunking, or semantic chunking.

Amazon Bedrock Agents offer you the ability to build and configure autonomous agents within your application. These agents securely connect to your company's data sources and augment user requests with the right information to generate accurate responses. You can create an agent in Amazon Bedrock with just a few quick steps, accelerating the time it takes to build generative AI applications. These agents support code interpretation to dynamically generate and execute code as well as return of control, which allows you to define an action schema and get the control back whenever the agent invokes the action. Additionally, Amazon Bedrock Agents can retain memory across interactions, offering more personalized and seamless user experiences.

Amazon Bedrock Flows is a workflow authoring and execution feature of Bedrock for generative AI applications. It accelerates the creation, testing, and deployment of user-defined generative AI workflows through an intuitive visual builder and a set of APIs. It allows you to seamlessly link the latest foundation models, Prompts, Agents, Knowledge Base, Guardrails, and AWS services (such as Amazon Lex, AWS Lambda, Amazon S3) along with business logic to build generative AI workflows. You can easily test and version your workflows, and run it in a secure serverless environment through a visual interface or API without having to stand up your own infrastructure.

Model Evaluation: With Amazon Bedrock model evaluation you pay for what you use, with no minimum volume commitments on the number of prompts or responses. For automatic (programmatic) evaluation, you only pay for the inference from your choice of model in the evaluation. The automatically-generated algorithmic scores are provided at no extra charge. During the Public Preview for automatic (Model/LLM-as-a-judge) evaluation, you only pay for the inference from your choice of generator model and evaluator model. In an LLM-as-a-judge model evaluation job, the built-in metrics use system judge prompt templates unique to each metric and available judge model that will be charged as part of your token usage, and the judge prompts are available in the public AWS documentation for transparency. For human-based evaluation where you bring your own workteam, you are charged for the model inference in the evaluation, and a charge of $0.21 per completed human task. A human task is defined as an instance of a human worker submitting an evaluation of a single prompt and its associated inference responses in the human evaluation user interface. The price is the same whether you have one or two models in your evaluation job and also the same regardless of how many evaluation metrics and rating methods you include. The charges for the human tasks will appear under the Amazon SageMaker section in your AWS bill and are the same for all AWS Regions. There is no separate charge for the workforce, as the workforce is supplied by you. For an evaluation managed by AWS, pricing is customized for your evaluation needs in a private engagement while working with the AWS expert evaluations team.

Amazon Bedrock Knowledge Bases Evaluation (RAG evaluation): With Amazon Bedrock Knowledge Bases Evaluation (RAG evaluation) you pay for what you use, with no minimum volume commitments on the number of prompts or responses. During the Public Preview, you only pay for the inference from your choice of generator model and evaluator model (the evaluation job uses an LLM-as-a-judge), as well as any charges incurred from using the Knowledge Base in the evaluation job according to Amazon Bedrock Knowledge Bases pricing. In a Knowledge Base Evaluation (RAG evaluation) job, the built-in metrics use system judge prompt templates unique to each metric and available judge model that will be charged as part of your token usage, and the judge prompts are available in the public AWS documentation for transparency. Some metrics involve doing judge model inference on retrieved context from your Knowledge Base or your ground truth answers in addition to the input prompt, which affects the costs associated with each metric - more information on each metric can be found in the public AWS documentation for evaluations.

Pricing Details

Pricing is dependent on the modality, provider, and model. Please select the model provider to see detailed pricing.

Amazon Bedrock offers select foundation models (FMs) from leading AI providers like Anthropic, Meta, Mistral AI, and Amazon for batch inference at a 50% lower price compared to on-demand inference pricing. Please refer to model list here.

  • AI21 Labs

  • AI21 Labs

    On-Demand pricing

    AI21 Labs models Price per 1,000 input tokens Price per 1,000 output tokens
    Jamba 1.5 Large $0.002 $0.008
    Jamba 1.5 Mini $0.0002 $0.0004
    Jurassic-2 Mid $0.0125 $0.0125
    Jurassic-2 Ultra $0.0188 $0.0188
    Jamba-Instruct $0.0005 $0.0007
  • Amazon

  • Amazon

    Other Amazon models Price 
    Amazon Rerank 1.0 $1.00 per 1,000 queries**
    **You are charged for number of queries where a query can contain up to 100 document chunks. If the query contains more than 100 document chunks, it is counted as multiple queries. For example, if a request contains 350 documents, it will be treated as 4 queries. Please note that each document can only contain upto 512 tokens (inclusive of the query and document’s total tokens), and if the token length is higher than 512 tokens, it is broken down into multiple documents.
  • Anthropic

  • Anthropic

    On-Demand and Batch pricing

    Region: US East (N. Virginia) and US West (Oregon)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3.5 Sonnet**

    $0.003

    $0.015

    $0.0015

    $0.0075

    Claude 3.5 Haiku

    $0.0008

    $0.004

    $0.0005

    $0.0025

    Claude 3 Opus*

    $0.015

    $0.075

    $0.0075

    $0.0375

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

    $0.000625

    Claude 3 Sonnet

    $0.003

    $0.015

    $0.0015

    $0.0075

    Claude 2.1

    $0.008

    $0.024

    N/A

    N/A

    Claude 2.0

    $0.008

    $0.024

    N/A

    N/A

    Claude Instant

    $0.0008

    $0.0024

    N/A

    N/A
    *Claude 3 Opus is currently available in the US West (Oregon) Region
    **Pricing for Claude 3.5 Sonnet is applicable to each version of Claude 3.5 Sonnet (v1 and v2) - Claude 3.5 Sonnet v2 is currently available in the US West (Oregon) Region

    Region: Europe (London)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3 Sonnet

    $0.003

    $0.015

    		 $0.0015 $0.0075

    Claude 3 Haiku

    $0.00025

    $0.00125

    		 $0.000125

    $0.000625

    Region: Europe (Zurich)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3.5 Sonnet

    $0.003

    $0.015

    $0.0015

    $0.0075

     

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

     

    Region: South America (Sao Paolo)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3 Sonnet

    $0.003

    $0.015

    $0.0015

    $0.0075

     

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

     

    Region: Canada (Central)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3 Sonnet

    $0.003

    $0.015

    $0.0015

     

    $0.0075

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

    Region: Asia Pacific (Mumbai)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3 Sonnet

    $0.003

    $0.015

    $0.0015

     

    $0.0075

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

    Region: Asia Pacific (Sydney)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3 Sonnet

    $0.003

    $0.015

    $0.0015

     

    $0.0075

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

    Region: Asia Pacific (Tokyo)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude Instant

    $0.0008

    $0.0024

    N/A

     

    N/A

    Claude 2.0/2.1

    $0.008

    $0.024

    N/A

    N/A

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

    Claude 3.5 Sonnet

    $0.003

    $0.015

    $0.0015

     

    $0.0075

    Region: Asia Pacific (Singapore)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude Instant

    $0.0008

    $0.0024

    $0.0004

     

    $0.0012

    Claude 2.0/2.1

    $0.008

    $0.024

    $0.004

     

    $0.012

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

    Claude 3.5 Sonnet

    $0.003

    $0.015

    N/A

    N/A

    Region: Europe (Paris)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

    Claude 3 Sonnet

    $0.003

    $0.015

    $0.0015

     

    $0.0075

    Region: Europe (Frankfurt)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude Instant

    $0.0008

    $0.0024

    N/A

    N/A

    Claude 2.0/2.1

    $0.008

    $0.024

    N/A

    N/A

    Claude 3 Sonnet

    $0.003

    $0.015

    $0.0015

     

    $0.0075

    Claude 3.5 Sonnet

    $0.003

    $0.015

    $0.0015

     

    $0.0075

    Claude 3 Haiku

    $0.00025

    $0.00125

    $0.000125

     

    $0.000625

    Region: Asia Pacific (Seoul)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3.5 Sonnet

    $0.003

    $0.015

    N/A

     

    N/A

    Claude 3 Haiku

    $0.00025

    $0.00125

    N/A

     

    N/A

    Region: US East (Ohio)

    Anthropic models Price per 1,000 input tokens Price per 1,000 output tokens Price per 1,000 input tokens (batch) Price per 1,000 output tokens (batch)

    Claude 3.5 Sonnet

    $0.003

    $0.015

    N/A

     

    N/A

    Claude 3 Haiku

    $0.00025

    $0.00125

    N/A

     

    N/A

    Latency Optimized Inference

    Region: US East (Ohio)

      Price per 1,000 input tokens Price per 1,000 output tokens
    Claude 3.5 Haiku $0.001 $0.005

    Provisioned Throughput pricing

    Region: US East (N. Virginia) and US West (Oregon)

    Anthropic models Price per hour per model with
    no commitment    		 Price per hour per model unit for 1-month commitment Price per hour per model unit for 6-month commitment

    Claude Instant

    		 $44.00

    $39.60

    $22.00

    Claude 2.0/2.1

    		 $70.00

    $63.00

    $35.00
    Anthropic models Price per hour per model with
    no commitment    		 Price per hour per model unit for 1-month commitment Price per hour per model unit for 6-month commitment

    Claude Instant

    		 $44.00

    $39.60

    $22.00

    Claude 2.0/2.1

    		 $70.00

    $63.00

    $35.00

    Region: Asia Pacific (Tokyo)

    Anthropic models Price per hour per model unit for 1-month commitment Price per hour per model unit for 6-month commitment

    Claude Instant

    $53.00

    $29.00

    Claude 2.0/2.1

    $86.00

    $48.00

    Region: Europe (Frankfurt)

    Anthropic models Price per hour per model unit for 1-month commitment Price per hour per model unit for 6-month commitment

    Claude Instant

    $49.00

    $27.00

    Claude 2.0/2.1

    $79.00

    $44.00

    Please reach out to your AWS account team for more details on model units. 

     

  • Cohere

  • Cohere

    On-Demand pricing

    Cohere models Price per 1,000 input tokens Price per 1,000 output tokens
    Command $0.0015 $0.0020
    Command-Light $0.0003 $0.0006
    Command R+ $0.0030 $0.0150
    Command R $0.0005 $0.0015
    Embed - English $0.0001 N/A
    Embed - Multilingual $0.0001 N/A
    Cohere models Price per 1,000 queries**
    Rerank 3.5 $2.00
    **You are charged for number of queries where a query can contain up to 100 document chunks. If the query contains more than 100 document chunks, it is counted as multiple queries. For example, if a request contains 350 documents, it will be treated as 4 queries. Please note that each document can only contain upto 500 tokens (inclusive of the query and document’s total tokens), and if the token length is higher than 512 tokens, it is broken down into multiple documents.

    Pricing for customization (fine-tuning)

    Cohere models Price to train 1,000 tokens Price to store each custom model per month Price to infer from a custom model per model unit per hour (with no-commit Provisioned Throughput pricing)

    Cohere Command

    $0.004

    $1.95

    $49.50
    Cohere Command-Light $0.001

    $1.95

    		 $8.56

    *Total tokens trained = number of tokens in training data corpus x number of epochs

    Provisioned Throughput pricing

    Cohere models Price per hour per model 
    with no commitment    		 Price per hour per model unit for 1-month commitment

    Price per hour per model unit for 6-month commitment

    Cohere Command

    		 $49.50

    $39.60

    $23.77
    Cohere Command - Light $8.56

    $6.85

    		 $4.11
    Embed - English $7.12

    $6.76

    		 $6.41
    Embed - Multilingual $7.12

    $6.76

    		 $6.41

    Please reach out to your AWS account or sales team for more details on model units. 

  • Meta Llama

  • Meta Llama

    Llama 3.2

    On-Demand and Batch pricing 

    Llama 3.1

    On-Demand and Batch pricing 

    Pricing for model customization (fine-tuning)

    Provisioned Throughput pricing

    Llama 3

    On-Demand pricing 

    Llama 2

    On-Demand pricing 

    Region: US East (N. Virginia) and US West (Oregon)

    Meta models Price per 1,000 input tokens Price per 1,000 output tokens

    Llama 2 Chat (13B)

    $0.00075

    $0.001
    Llama 2 Chat (70B)

    $0.00195

    		 $0.00256

    Pricing for model customization (fine-tuning)

    Meta models Price to train 1,000 tokens Price to store each custom model* per month Price to infer from a custom model for 1 model unit per hour (with no-commit Provisioned Throughput pricing)

    Llama 2 Pretrained (13B)

    $0.00149

    $1.95

    $23.50
    Llama 2 Pretrained (70B)

    $0.00799

    		 $1.95 $23.50

    *Custom model storage = $1.95

    Provisioned Throughput pricing

    Meta models Price per hour per model unit for 1-month commitment Price per hour per model unit for 6-month commitment

    Llama 2 Pretrained and Chat (13B)

    $21.18

    $13.08
    Llama 2 Pretrained (70B)

    $21.18

    		 $13.08

    *Llama 2 Pre-trained models are available only in provisioned throughput after customization. 

    Please reach out to your AWS account or sales team for more details on model units.

  • Mistral AI

  • Mistral AI

  • Stability AI

  • Stability AI

    On-Demand pricing

    Stability AI Model Price per generated image
    Stable Image Core $0.04
    SD3 Large $0.08
    Stable Image Ultra $0.14

    Previous generation of image models offered by Stability AI are priced per image, depending on step count and image resolution.

    Stability AI model Image resolution Price per image generated for standard quality (<=50 steps) Price per image generated for premium quality (>50 steps)
    SDXL 1.0 Up to 1024 x 1024 $0.04 $0.08

    Provisioned Throughput pricing

    Stability AI model Price per hour per model unit for 1-month commitment* Price per hour per model unit for 6-month commitment*

    SDXL 1.0

    $49.86

    $46.18

    *Includes inference for base and custom models

    Please reach out to your AWS account or sales team for more details on model units.

    Currently, model customization (fine-tuning) is not supported for Stability AI models on Amazon Bedrock.

  • Custom Model Import

  • Custom Model Import

    • Llama

    • Regions: US East (N. Virginia) and US West (Oregon)

      Custom Model Unit version v1.0
      Price per Custom Model Unit per min* $0.0785
      Monthly storage cost per Custom Model Unit $1.95
       
      The Custom Model Units needed to host a model depend on a variety of factors - notably the model architecture, model parameter count, and context length. The exact number of Custom Model Units needed will be determined at the time of import. For reference, Llama 3.1 8B 128K model requires 2 Custom Model Units, a Llama 3.1 70B 128k model requires 8 Custom Model Units.
      *Billed in 5 minute windows
    • Multimodal Llama

    • Regions: US East (N. Virginia) and US West (Oregon)

      Custom Model Unit version v1.0
      Price per Custom Model Unit per min* $0.0785
      Monthly storage cost per Custom Model Unit $1.95
       
      The Custom Model Units needed to host a model depend on a variety of factors - notably the model architecture, model parameter count, and context length. The exact number of Custom Model Units needed will be determined at the time of import. For reference, Llama 3.2 11B 128K model requires 4 Custom Model Units.
      *Billed in 5 minute windows
    • Mistral

    • Regions: US East (N. Virginia) and US West (Oregon)

      Custom Model Unit version v1.0
      Price per Custom Model Unit per min* $0.0785
      Monthly storage cost per Custom Model Unit $1.95
       
      The Custom Model Units needed to host a model depend on a variety of factors - notably the model architecture, model parameter count, and context length. The exact number of Custom Model Units needed will be determined at the time of import. For reference, Mistral 7B 32K model requires 1 Custom Model Unit.
      *Billed in 5 minute windows
    • Mixtral

    • Regions: US East (N. Virginia) and US West (Oregon)

      Custom Model Unit version v1.0
      Price per Custom Model Unit per min* $0.0785
      Monthly storage cost per Custom Model Unit $1.95
       
      The Custom Model Units needed to host a model depend on a variety of factors - notably the model architecture, model parameter count, and context length. The exact number of Custom Model Units needed will be determined at the time of import. For reference, Mixtral 8x7B 32K model requires 4 Custom Model Units.
      *Billed in 5 minute windows
    • Flan

    • Regions: US East (N. Virginia) and US West (Oregon)

      Custom Model Unit version v1.0
      Price per Custom Model Unit per min* $0.0785
      Monthly storage cost per Custom Model Unit $1.95
       
      The Custom Model Units needed to host a model depend on a variety of factors - notably the model architecture, model parameter count, and context length. The exact number of Custom Model Units needed will be determined at the time of import. For reference, Flan-T5 XL 512 model requires 1 Custom Model Unit. 
      *Billed in 5 minute windows

    On-Demand Inference Pricing:
    You are billed in 5-minute windows for the duration your model copy is active starting from the first successful invocation. The maximum throughput and concurrency limit per model copy depends on factors such as input/output token mix, hardware type, model size, architecture, inference optimizations, and is determined during the model import workflow.

    Bedrock automatically scales the number of model copies depending on your usage patterns. If there are no invocations for a 5-minute period, Bedrock will scale down to zero and scale back up when you invoke your model. While scaling back up, you may experience a cold-start duration (in tens of seconds) depending on model size. Bedrock also scales up the number of model copies if your inference volume consistently exceeds the concurrency limits of a single model copy. Note: There is a default maximum of 3 model copies per account per imported model that can be increased through Service Quotas.

Pricing Tools (Details)

  • Flows

  • Amazon Bedrock Flows

    You are charged based on the number of node transitions required to execute your application. Bedrock Flows counts a node transition each time a node in your workflow is executed. You are charged for the total number of node transitions across all your flows.

    All charges are metered daily and billed monthly starting February 1st, 2025.

    Price per 1,000 node transitions
    $0.035

    Additional Charges

    You may incur additional charges if the execution of your application workflow utilizes other AWS services or transfers data. For example, if your workflow invokes an Amazon Bedrock Guardrail policy, you will be billed for the number of text units processed by the policy. 

  • Knowledge Bases

  • Rerank models

    Rerank models are designed to improve the relevance and accuracy of responses in Retrieval Augmented Generation (RAG) applications. They are charged per request.

    **You are charged for number of queries where a query can contain up to 100 document chunks. If the query contains more than 100 document chunks, it is counted as multiple queries. For example, if a request contains 350 documents, it will be treated as 4 queries. Please note that each document can only contain upto 512 tokens (inclusive of the query and document’s total tokens), and if the token length is higher than 512 tokens, it is broken down into multiple documents. A query is equivalent to a search unit.

  • Amazon Bedrock Guardrails

  • Amazon Bedrock Guardrails

    Guardrail policy*

    Price per 1,000 text units**

    Content filters

    $0.75

    Denied topics

    $1

    Contextual grounding check***

    $0.1

    Sensitive information filter (PII)

    $0.1

    Sensitive information filter (regular expression) 

    Free

    Word filters

    Free

    On-Demand pricing

    * Each guardrail policy is optional and can be enabled based on your application requirements. Charges will be incurred based on the policy type used in the guardrail. For example, if a guardrail is configured with content filters and denied topics, charges will be incurred for these two policies, while there will be no charges associated with sensitive information filters.

    **A text unit can contain up to 1000 characters. If a text input is more than 1000 characters, it is processed as multiple text units, each containing 1000 characters or less. For example, if a text input contains 5600 characters, it will be charged for 6 text units.

    *** Contextual grounding check uses a reference source and a query to determine if the model response is grounded based on the source and relevant to the query. The total number of text units charged is calculated by combining all the characters in the source, query, and model response. 

    Guardrails are not supported for images and embeddings.

  • Model Evaluation

  • Model Evaluation

    Model evaluation is charged for the inference from your choice of model. Automatically-generated algorithmic scores are provided at no extra charge. For human-based evaluation where you bring your own workstream, you are charged for the model inference in the evaluation, and a charge of $0.21 per completed human task.

    Model

    Price per 1,000 input tokens

    Price per 1,000 output tokens

    Price per human task

    		  

    Model selected for evaluation

    Based on model selected

    Based on model selected

     

    $0.21

    		  

Pricing examples

  • An application developer makes the following API calls to Amazon Bedrock: a request to AI21’s Jurassic-2 Mid model to summarize an input of 10K tokens of input text to an output of 2K tokens.

    Total cost incurred = 10K tokens/1000 * $0.0125 + 2K tokens/1000 * $0.0125 = $0.15

  • On-Demand pricing

    An application developer makes the following API calls to Amazon Bedrock on an hourly basis: a request to Amazon Titan Text Lite model to summarize an input of 2K tokens of input text to an output of 1K tokens.

    Total hourly cost incurred is = 2K tokens/1000 * $0.0003 + 1K tokens/1000 * $0.0004 = $0.001.

    An application developer makes the following API calls to Amazon Bedrock: a request to the Amazon Titan Image Generator base model to generate 1000 images of 1024 x 1024 in size of standard quality.

    Total cost incurred = 1000 images * $0.01 per image = $10

    Customization (fine-tuning and continued pretraining) pricing

    An application developer customizes an Amazon Titan Image Generator model using 1000 image-text pairs. After training, the developer uses custom model provisioned throughput for 1 hour to evaluate the performance of the model. The fine-tuned model is stored for 1 month. After evaluation, the developer uses provisioned throughput (1-month commitment term) to host the customized model.

    Monthly cost incurred for fine-tuning = fine-tuning training ($.005 * 500 * 64), where $0.005 is the price per image seen, 500 is the number of steps, and 64 is the batch size, + custom model storage per month ($1.95) + 1 hour of custom model inference ($21) = $160 + $1.95 + 21 = $182.95

    Provisioned Throughput pricing

    An application developer buys two model units of Amazon Titan Text Express with a 1-month commitment for their text summarization use case.

    Total monthly cost incurred = 2 model units * $18.40/hour * 24 hours * 31 days = $27,379.20

    An application developer buys one model unit of the base Amazon Titan Image Generator model with a 1-month commitment.

    Total cost incurred = 1 model unit * $16.20 * 24 hours * 31 days = $12,052.80

  • On-Demand pricing

    An application developer makes the following API calls to Amazon Bedrock in the US West (Oregon) Region: a request to Anthropic’s Claude model to summarize an input of 11K tokens of input text to an output of 4K tokens.

    Total cost incurred = 11K tokens/1000 * $0.008 + 4K tokens/1000 * $0.024 = $0.088 + $0.096 = $0.184

    Provisioned Throughput pricing

    An application developer buys one model unit of Anthropic Claude Instant in the US West (Oregon) Region:

    Total monthly cost incurred = 1 model unit * $39.60 * 24 hours * 31 days = $29,462.40

  • On-Demand pricing

    An application developer makes the following API calls to Amazon Bedrock: a request to Cohere’s Command model to summarize an input of 6K tokens of input text to an output of 2K tokens.

    Total cost incurred = 6K tokens/1,000 * $0.0015 + 2K tokens/1,000 * $0.0020 = $0.013

    An application developer makes the following API calls to Amazon Bedrock: A request to Cohere’s Command - Light model to summarize an input of 6K tokens of input text to an output of 2K tokens.

    Total cost incurred = 6K tokens/1000 * $0.0003 + 2K tokens/1000 * $0.0006 = $0.003

    An application developer makes the following API calls to Amazon Bedrock: A request to either Cohere’s Embed English or Embed Multilingual model to generate embeddings for 10K tokens of input.

    Total cost incurred = 10K tokens/1000 * $0.0001 = $.001

    Customization (fine-tuning) pricing

    An application developer customizes a Cohere Command model using 1000 tokens of data. After training, uses custom model provisioned throughput for 1 hour to evaluate the performance of the model. The fine-tuned model is stored for 1 month. After evaluation, the developer uses provisioned throughput (1mo commit) to host the customized model.

    Monthly cost incurred for fine-tuning = Fine-tuning training ($0.004 * 1000) + custom model storage per month ($1.95) + 1 hour of custom model inference ($49.50) = $55.45

    Monthly cost incurred for provisioned throughput (1-month commitment) of custom model = $39.60

    Provisioned Throughput pricing

    An application developer, buys one model unit of Cohere Command with a 1-month commitment for their text summarization use case.

    Total monthly cost incurred = 1 model unit * $39.60 * 24 hours * 31 days = $29,462.40

  • On-Demand pricing

    An application developer makes the following API calls to Amazon Bedrock: a request to Meta’s Llama 2 Chat (13B) model to summarize an input of 2K tokens of input text to an output of 500 tokens.

    Total cost incurred = 2K tokens/1000 * $0.00075 + 500 tokens/1000 * $0.001 = $0.002

    Customization (fine-tuning) pricing

    An application developer customizes the Llama 2 Pretrained (70B) model using 1000 tokens of data. After training, uses custom model provisioned throughput for 1 hour to evaluate the performance of the model. The fine-tuned model is stored for 1 month. After evaluation, the developer uses provisioned throughput (1mo commit) to host the customized model.

    Monthly cost incurred for fine-tuning = Fine tuning training ($0.00799 * 1000) + custom model storage per month ($1.95) + 1 hour of custom model inference ($23.50) = $33.44

    Monthly cost incurred for provisioned throughput (a 1-month commit) of custom model = $21.18

    Provisioned Throughput pricing

    An application developer buys one model unit of Meta Llama 2 with a 1-month commitment for their text summarization use case.

    Total monthly cost incurred = 1 model unit * $21.18 * 24 hours * 31 days = $15,757.92

  • On-Demand pricing

    An application developer makes the following API calls to Amazon Bedrock on an hourly basis: a request to Mistral 7B model to summarize an input of 2K tokens of input text to an output of 1K tokens.

    Total hourly cost incurred = 2K tokens/1000 * $0.00015 + 1K tokens/1000 * $0.0002 = $0.0005

    An application developer makes the following API calls to Amazon Bedrock on an hourly basis: a request to Mixtral 8x7B model to summarize an input of 2K tokens of input text to an output of 1K tokens.

    Total hourly cost incurred = 2K tokens/1000 * $0.00045 + 1K tokens/1000 * $0.0007 = $0.0016

    An application developer makes the following API calls to Amazon Bedrock on an hourly basis: a request to Mistral Large model to summarize an input of 2K tokens of input text to an output of 1K tokens. 

    Total hourly cost incurred = 2K tokens/1000 * $0.008 + 1K tokens/1000 * $0.024 = $0.04

  • On-Demand pricing

    An application developer makes the following API calls to Amazon Bedrock: a request to the SDXL model to generate a 512 x 512 image with a step size of 70 (premium quality).

    Total cost incurred = 1 image * $0.036 per image = $0.036

    An application developer makes the following API calls to Amazon Bedrock: A request to the SDXL 1.0 model to generate a 1024 x 1024 image with a step size of 70 (premium quality).

    Total cost incurred = 1 image * $0.08 per image = $0.08

    Provisioned Throughput pricing

    An application developer buys one model unit of SDXL 1.0 with a 1-month commitment.

    Total cost incurred = 1 * $49.86 * 24 hours * 31 days = $37,095.84

  • Model evaluation example 1:

    On-demand pricing
    An application developer submits a dataset for human-based model evaluation using Anthropic Claude 2.1 and Anthropic Claude Instant in the US East (N. Virginia) AWS Region.

    The dataset contains 50 prompts, and the developer requires one worker to rate each prompt-response set (configurable in the evaluation job creation as “workers per prompt” parameter).

    There will be 50 tasks in this evaluation job (one task for each prompt-response set per each worker). The 50 prompts combine to 5000 input tokens, and the associated responses combine to 15,000 tokens for Anthropic Claude Instant and 20,000 tokens for Anthropic Claude 2.1.

    The following charges are incurred for this model evaluation job:

    Item Number of input tokens Price per 1000 input tokens Cost of input Number of output tokens Price per 1000 output tokens Cost of output Number of human tasks Price per human task Cost of human tasks Total
    Claude Instant Inference 5000 $0.0008 $0.004 15000 $0.0024 $0.036       $0.04
    Claude 2.1 Inference 5000 $0.008 $0.04 20000 $0.024 $0.48       $0.52
    Human Tasks             50 $0.21 $10.50 $10.50
    Total                   $11.06

    Model evaluation example 2:

    On-demand pricing
    An application developer submits a dataset for human-based model evaluation using Anthropic Claude 2.1 and Anthropic Claude Instant in the US East (N. Virginia) AWS Region.

    The dataset contains 50 prompts, and the developer requires two workers to rate each prompt-response set (configurable in the evaluation job creation as “workers per prompt” parameter). There will be 100 tasks in this evaluation job (1 task for each prompt-response set per each worker: 2 workers x 50 prompt-response sets = 100 human tasks).

    The 50 prompts combine to 5000 input tokens, and the associated responses combine to 15000 tokens for Anthropic Claude Instant and 20000 tokens for Anthropic Claude 2.1.

    The following charges are incurred for this model evaluation job:

    Item Number of input tokens Price per 1000 input tokens Cost of input Number of output tokens Price per 1000 output tokens Cost of output Number of human tasks Price per human task Cost of human tasks Total
    Claude Instant Inference 5000 $0.0008 $0.0040 15000 $0.0024 $0.036       $0.04
    Claude 2.1 Inference 5000 $0.008 $0.0400 20000 $0.024 $0.48       $0.52
    Human Tasks             100 $0.21 $21.00 $21.00
    Total                   $21.56

  • Example 1: Customer support chatbot
    An application developer creates a customer support chatbot and uses content filters to block harmful content and denied topics to filter undesirable queries and responses.

    The chatbot serves 1000 user queries per hour. Each user query has an average input length of 200 characters and receives a FM response of 1500 characters.

    Each user query of 200 characters correspond to 1 text unit.

    Each FM response of 1,500 characters correspond to 2 text units.

    Text units processed each hour = (1 + 2) * 1000 queries = 3000 text units

    Total cost incurred per hour for content filters and denied topic = 3000 * ($0.75 + $1.00) / 1000 = $5.25

     

    Example 2: Call center transcript summarization
    An application developer creates an application to summarize chat transcripts between users and support agents. It uses sensitive information filter to redact personally identifiable information (PII) in the generated summaries for 10,000 conversations.

    Each generated summary has an average of 3,500 characters that corresponds to 4 text units.

    Total cost incurred to summarize 10,000 conversations = 10000 * 4 * ($0.1/1000) = $4

    Item Number of input tokens Price per 1000 input tokens Cost of input Number of output tokens Price per 1000 output tokens Cost of output Number of human tasks Price per human task Cost of human tasks Total
    Claude Instant Inference 5000 $0.0008 $0.004 15000 $0.0024 $0.036       $0.04
    Claude 2.1 Inference 5000 $0.008 $0.04 20000 $0.024 $0.48       $0.52
    Human Tasks             100 $0.21 $21.00 $21.00
    Total                   $21.56

  • Pricing Example: An application developer imports a customized Llama 3.1 type model that is 8B parameter in size with a 128K sequence length in us-east-1 region and deletes the model after 1 month. This requires 2 Custom Model Units. So, the price per minute will be $0.1570 because 2 Custom Model Units are required. The model storage costs for 2 Custom Model Units would be $3.90 for the month.

    There is no charge to import the model. The first successful invocation is at 8:03 AM, at which time the metering starts. The 5-minute metering windows are from 8:03 AM - 8:07 AM; 8:07 AM - 8:11 AM, and so on. If there is at least one invocation during any 5-minute period, the window will be considered active for billing. If there is no invocation from 8:07 AM - 8:11 AM, the metering will stop at 8:11 AM. In this case, the bill would be calculated as follows: $0.1570 * 5 minutes * 3 five minute windows = $2.355.

  • Pricing Example 1 (Reranking using Amazon Rerank 1.0 model)

    In a given month, you make 2 million requests to Rerank API using Amazon Rerank 1.0 model – 1 million requests contain fewer than 100 documents each and hence will be charged for one request each. The remaining 1 million requests contain 120-150 documents, and hence each request will be charged for 2 requests.

    Price for one request = $0.001
    Total charge = 1,000,000 * $0.001 + 1,000,000*2*$0.001= $3000 

  • Example: News summarization
    An application developer creates a flow to automate news summarization for traders. The flow includes an Input node that takes in an S3 location, and a S3 retrieval node that retrieves 10 files that include articles from 10 major news agency in S3 (2 node transitions). It then uses an iterator node to invoke a model with a prompt node to summarize each file (+ 10 files x 2 node transitions). It then collects all the results using a collector node, write the results to S3 using S3 storage node, and complete in an Output node (+ 3 node transition). They run this flow every half hour of every week day.

    The number of node transition per flow execution is: 2+1+10*2 + 3 = 25 node transitions/flow execution

    The number of flow execution per month is: 24 hours *2* 5 days * 4 weeks = 960 flow executions/month.

    Total monthly bill is: 25 * 960 * $0.035/1000 = $0.84

    Additional charges
    The bill will also include additional charges for AWS services used in the workflow execution, including Amazon S3 usages in the retrieval and storage nodes, and Amazon Bedrock foundation model usage in the prompt node.