Amazon Bedrock Pricing: Model Rates and a $1,270 App Budget

Amazon Bedrock pricing matches Anthropic's standard API rate for Claude Sonnet 5.5 when you use global routing: $2 per million input tokens and $10 per million output tokens. An AWS Bedrock pricing estimate still needs retrieval, storage and safeguards: the support feature below totals $1,270 a month for 100,000 answers. Start with on-demand inference, then buy caching, batch processing or committed capacity only where your workload qualifies.

Verified against AWS's live pricing page, October 2026, with the model tables and direct-provider comparisons checked on 19 October 2026. Prices are USD public list rates before tax, credits and negotiated discounts. The workload examples are calculations, not production measurements.

Amazon Bedrock Pricing at a Glance

Amazon Bedrock, Amazon Web Services' managed service for running foundation models, charges for the model and the surrounding services you use. A foundation model is the underlying AI that reads your request and generates the answer. There is no single Bedrock token price or mandatory monthly seat subscription.

Amazon Bedrock Model Pricing

This shortlist covers Claude, Amazon Nova, Meta Llama and Mistral AI models a startup might evaluate for an AWS feature. It uses Standard on-demand text rates, with US East (N. Virginia) as the source Region. “Global” means global cross-region inference; “Geo/in-region” means the corresponding AWS pricing section. Check your selected model's availability and regional row before deploying. Every rate comes from AWS's Bedrock pricing page.

Haiku 5.5's long-context rate applies to a request whose total prompt exceeds 100,000 tokens, including cached input. It is a different price band, not a surcharge on just the excess tokens. Keep that boundary in a document-processing estimate. Anthropic pricing explains the threshold; AWS lists both price rows.

Choose the least expensive model that passes your feature's acceptance criteria. A routing classifier and a difficult customer dispute need different evaluations. These prices alone cannot establish answer quality, retry frequency or whether a cheaper model completes the same task.

Bedrock Token Pricing

Input and output are separate meters. Tokens are the text units a model processes. Input includes the material you send, such as instructions, conversation history and retrieved documents. Output is the model's generated material, including any reasoning tokens billed as output for the model you select.

For an uncached text request, the calculation is:

Model cost = input tokens ÷ 1,000,000 × input rate + output tokens ÷ 1,000,000 × output rate.

At Sonnet 5.5's global rates, 3,000 input tokens and 500 output tokens cost $0.011 per answer, or $11 per 1,000 answers. A conversation can contain several model calls. Count all of them, including tool loops and retries; counting customers or chat sessions alone understates usage.

The practical wall is context growth. If your support assistant sends a longer conversation and more documentation on every turn, its input bill grows even when the number of users stays flat. Track tokens per completed answer alongside request volume.

Which Bedrock Billing Mode Fits Your Workload?

Standard on-demand is the sensible starting point for a new interactive feature. It keeps inference spend tied to usage while you learn your traffic pattern. AWS also offers Flex, Priority, Batch, Reserved capacity and model-specific Provisioned Throughput. These options have different availability and commitment rules. AWS service tiers describes the operating differences.

Batch suits work that can finish later. You submit a collection of requests and retrieve the results asynchronously. AWS advertises a 50% discount for selected models. For example, Nova 2 Lite global Batch is $0.15 input and $1.25 output per million tokens; Claude Sonnet 4.6 global Batch is $1.50 and $7.50. Check the actual model row, not just the headline discount. AWS pricing

Flex suits a single call that can tolerate a slower response. It remains synchronous, so your application waits for the result, unlike a submitted batch job. Nova 2 Lite global Flex has the same $0.15/$1.25 rates as its Batch row. A background classification task is a better fit than a customer waiting at checkout.

Priority buys preferential processing. For Nova 2 Lite global, AWS lists $0.525/$4.375, versus Standard's $0.30/$2.50. Pay that premium when response speed has a measured business benefit. An overnight document export gains little from it. These are Nova's published rates, not a promise that every model has the same tier menu. AWS pricing

AWS Bedrock Provisioned Throughput Pricing

Provisioned Throughput charges for reserved model capacity by the hour, so unused capacity still has a cost. AWS's Nova table lists $60.50 per model-unit-hour without commitment, $55 with a one-month commitment, and $30.25 with a six-month commitment. A model unit is AWS's capacity allocation for that model; its useful throughput depends on the model and workload. Anthropic Provisioned Throughput pricing requires an account-team quote. AWS pricing

For a 30-day month, a single Nova unit at the one-month rate costs 720 × $55 = $39,600. At the six-month rate, that month's portion is $21,780, with the remaining commitment still to pay. A lower hourly rate does not make that commitment appropriate for a startup serving intermittent requests.

Reserved tier is a separate capacity meter. It reserves input and output tokens per minute, abbreviated TPM, with one-month or three-month commitments. For Sonnet 4.6 global, AWS lists $0.18 per hour per 1,000 input TPM and $0.90 per hour per 1,000 output TPM with a one-month term. The three-month rates are $0.162 and $0.81. AWS pricing

The distinction matters when calculating a break-even. At full utilization, 1,000 tokens per minute produces 60,000 tokens per hour. Sonnet 4.6 Standard charges $0.18 for that input volume or $0.90 for that output volume. The three-month Reserved rates therefore break even at 90% sustained utilization of each reserved dimension, before caching or batch savings. This is a rate comparison, not a statement about minimum purchasable capacity or your contract terms.

Get a quote after measuring steady throughput and the input/output mix. Bursty demand can leave one reserved dimension underused while the other reaches its limit. There is no annual subscription discount to chase here; the financial risk is paying for committed capacity your application does not consume.

Prompt Caching: Price the Write and the Reuse

Caching saves money when the same prompt prefix is reused. A prefix is the unchanged beginning of a request, such as a policy document or instructions. AWS's prompt-caching documentation specifies model-dependent minimum lengths, supported fields and cache lifetimes. For Sonnet 5.5, the minimum is 512 tokens per checkpoint, with five-minute and one-hour lifetimes supported.

Sonnet 5.5 global costs $2.50 per million tokens to write a five-minute cache, $4 to write a one-hour cache, and $0.10 to read it. Ordinary uncached input remains $2. AWS explicitly dates a 50% cache-read price reduction to October 7, 2026. AWS pricing

Suppose ten calls reuse the same 10,000-token prefix within the cache lifetime. Without caching, those prefix tokens cost $0.20. With one five-minute write and nine hits, they cost $0.025 + $0.009 = $0.034, an 83% saving on the prefix. New input and generated output remain additional charges.

For this model, a five-minute write pays back after one reuse. A one-hour write needs two reuses: the longer lifetime costs more to establish. A timestamp or changing customer field placed inside the supposedly stable prefix can undermine that saving. Put stable material first and watch actual cache-read usage before promising a lower bill.

Cross-Region Inference: Routing Is Free, Geography Can Change the Rate

Cross-region inference has no additional routing fee, but the chosen inference geography can change token prices. Geographic routing keeps processing within the profile's geography; global routing can send it to supported Regions worldwide. AWS calculates prices from the source Region. AWS cross-region documentation

For Sonnet 5.5, the global row is $2/$10, while the Geo/in-region row is $2.20/$11 per million input/output tokens. Nova 2 Lite likewise changes from $0.30/$2.50 globally to $0.33/$2.75 in the Geo/in-region row. Those examples have a 10% premium over global. AWS pricing

For a product with geographic processing requirements, use the eligible profile and budget its rate. For a product without that constraint, global routing can reduce inference spend. A source Region in the United States does not by itself mean a global profile processes every request there.

Bedrock Versus the Model Maker's Own API

Compare the same model, geography, service tier and cache behavior. A base-rate tie can disappear when one route supports a discount that the other does not. Keep the full application cost separate from this inference comparison.

AWS Bedrock Claude Pricing

Anthropic's direct Claude API and Bedrock global Standard currently tie for Sonnet 5.5 at $2/$10, Opus 5.5 at $4/$20, and short-context Haiku 5.5 at $0.10/$0.50. Sonnet 5.5's published cache-write and cache-read rates also match. Anthropic pricing, AWS pricing

For the example workload of 300 million input and 50 million output tokens, Sonnet 5.5 inference costs $1,100 through either global Standard route. Bedrock Geo/in-region costs $1,210. However, Anthropic's direct US-only inference also applies a 1.1× multiplier, making that comparable direct route $1,210 too. Geographic processing is not a Bedrock-only premium.

The larger difference is asynchronous work. Anthropic lists Sonnet 5.5 Batch at $1/$5, making that same token volume $550. AWS's Sonnet 5.5 batch cells currently say N/A. A team that can wait for results has a concrete reason to compare the direct Batch API before choosing Bedrock Standard.

Bedrock is a strong default when your feature already runs on AWS and the eligible rate matches. Choose the direct API when its available discount or required feature materially improves your application. Price the integration work too; moving a working service solely to reproduce the same token rate does not create savings.

Mistral Direct: A US Price Tie, a Regional Difference

Mistral AI lists Mistral Large 3 at $0.50 input and $1.50 output per million tokens on its own API pricing page. AWS lists the same Standard rates in Virginia, Ohio and Oregon. The example's 300 million input and 50 million output tokens cost $225 through either of those routes.

AWS's Mumbai row is $0.59/$1.76, making the same token workload $265, or $40 more than direct Standard. That is a comparison with the maker's standard offer; match any required geography before treating it as an equivalent deployment. Mistral's direct page also lists $0.05 cached input for Large 3, so repeated-prompt traffic deserves a separate estimate. AWS pricing, Mistral pricing

For Meta Llama and Amazon Nova, use the AWS rows to price the Bedrock deployment. A reseller's Llama price is not a direct Meta API quote. Within this shortlist, the same illustrative token volume costs $84 on Llama 4 Scout or $215 on Nova 2 Lite global. Run your acceptance evaluation before interpreting those amounts as substitutes for Claude's answers.

For a broader provider shortlist, the site's cheapest AI API comparison adds other hosting choices. Keep each provider's own prices and deployment conditions attached to the calculation.

Knowledge Bases, Guardrails and Agents: The Additional Meters

Optional services can outweigh model tokens on a small-model workload. Retrieval-augmented generation, or RAG, means finding relevant material and adding it to a model's prompt. That introduces storage and retrieval work before the model generates anything.

Amazon Bedrock Knowledge Bases Pricing

Amazon Bedrock now lists both Managed Knowledge Bases and Self-Managed Knowledge Bases. For the managed offering, storage is $5 per GB of raw data per month, and standard retrieval is $1 per 1,000 Retrieve API calls. The managed parser, embedding model and reranker are included. Embeddings convert content into searchable numerical representations; reranking reorders retrieved passages by relevance. Selecting your own embedding or reranking model adds model charges. AWS pricing

Managed agentic retrieval adds a planning step. With the managed planning model, it costs $4 per 1,000 Agentic Retrieve calls plus $1 per 1,000 underlying Retrieve calls. At 100,000 agentic calls with two underlying retrieves each, that is $600 in retrieval charges, compared with $100 for 100,000 standard retrieves. Storage and answer-generation tokens remain additional.

Self-managed retrieval has a different budget: your embedding model and backing vector store have their own charges. AWS's application cost guide names those components. Do not copy a fixed vector-database minimum into every Bedrock estimate; the managed and self-managed architectures have different meters.

Other published options include $2 per 1,000 SQL-generation requests, Cohere Rerank 3.5 at $2 per 1,000 billed queries, and Bedrock Data Automation Standard Output at $0.010 per document page when used as a parser. A rerank query covers up to 100 document chunks, subject to the model's chunk-length rules; larger requests can count as multiple queries. Choose these options deliberately instead of adding them automatically to the managed offering's included components. AWS pricing

Amazon Bedrock Guardrails Pricing

Amazon Bedrock Guardrails bills each enabled filter separately. Its standard/classic table lists $0.15 per 1,000 text units for content filters and another $0.15 for denied topics. Sensitive-information filters and contextual-grounding checks each cost $0.10 per 1,000 text units; automated-reasoning checks cost $0.17 per 1,000 units per policy. Word filters and regular-expression sensitive-information filters are free. AWS pricing

A text unit is up to 1,000 characters, not 1,000 model tokens. Each assessed input or output rounds up. Grounding checks count the reference source, query and response, so sending an entire document through that check changes the calculation.

AWS also lists a separate InvokeGuardrailChecks API rate card: $0.07 for content, $0.08 for prompt attack, and $0.10 for sensitive-information checks per 1,000 text units. Use the price table for the API you invoke. A lower content-filter number from that table is not a blanket discount on the standard integration.

Amazon Bedrock AgentCore Pricing

Amazon Bedrock AgentCore, the service for deploying and operating agents, has its own consumption charges. AWS says Agents Classic is no longer available to new customers and directs new builders toward AgentCore. An older “agents cost only tokens” estimate is therefore an incomplete starting point for a new implementation. AWS Agents notice

On the AgentCore pricing page, Runtime microVM v1 costs $0.0895 per vCPU-hour plus $0.00945 per GB-hour of memory. Runtime v2 consumption costs $0.1276 per vCPU-hour plus $0.0169 per GB-hour. CPU and memory are separate measurements. CPU waiting for an external call can consume no billable CPU, while memory can still incur charges.

Other agent components add their own meters. Gateway API invocations are $0.005 per 1,000, and Gateway search calls are $0.025 per 1,000. Since October 6, 2026, short-term Memory is priced at $1 per GB ingested, $0.20 per GB retrieved and $0.10 per GB-month stored. Browser, code execution, long-term memory and observability also require their own usage estimates. These costs sit alongside the model calls your agent makes.

For existing Agents Classic applications, budget the underlying models, retrieval and actions. An agent that looks up an order, calls a business API and drafts an answer can make several model calls per completed task. The inference-platforms guide helps frame that cost per completed task.

Other Costs to Keep in the Estimate

AWS lists Intelligent Prompt Routing at $1 per 1,000 requests, Flows at $0.035 per 1,000 node transitions, and the Simple Prompt Optimizer at $0.03 per 1,000 input and optimized-output tokens. Human model evaluation adds $0.21 per completed human task, plus inference. Fine-tuning, custom-model storage and deployment introduce separate charges. AWS pricing

For a production feature, also estimate application compute, source storage, logs and networking. A long output, repeated retrieval, a failed call retried successfully, and a background evaluation all consume resources without necessarily creating another customer-visible answer.

AWS Bedrock Pricing Calculator: A Worked Monthly Bill

A bounded support feature can cost $1,270 a month in the Bedrock components modeled here. Assume an existing SaaS application answers 100,000 support questions each month using Sonnet 5.5 global Standard. Each completed answer uses one model call with 3,000 billed input tokens and 500 billed output tokens. There are no retries, cache hits or batch discounts in this baseline.

The feature stores 5 GB of raw documentation in a Managed Knowledge Base and makes one standard retrieval per answer. Content filtering checks a 500-character user message and a 2,000-character final response. Those character counts are separate illustrative assumptions from the model-token counts. The application does not use AgentCore.

• Model input: $600. 100,000 × 3,000 = 300 million tokens, at $2 per million.
• Model output: $500. 100,000 × 500 = 50 million tokens, at $10 per million.
• Managed knowledge-base storage: $25. 5 GB × $5 per month.
• Standard retrieval: $100. 100,000 calls ÷ 1,000 × $1.
• Content filtering: $45. One input text unit plus two output units per answer gives 300,000 units, at $0.15 per 1,000.

Monthly modeled subtotal: $1,270, or $0.0127 per answer. That is $12.70 per 1,000 answers, using the rates on AWS's pricing page.

Changing the model to short-context Haiku 5.5 reduces the same token bill to $55 and the modeled subtotal to $225, if it passes your quality requirements. Retrieval, storage and content filtering still cost $170. Once model tokens get cheap, those surrounding services become the main optimization target.

Response length also matters. Doubling Sonnet's output from 500 to 1,000 tokens adds $500 per month. Doubling input from 3,000 to 6,000 tokens adds $600. Measure those quantities before negotiating a small token discount.

Amazon Bedrock Pricing Calculator: Put Your Own Workload Into It

Use the AWS Pricing Calculator for an architecture estimate, then retain a calculation you can reconcile with billing. An estimate without model, Region and token assumptions is difficult to update.

1. 1

   Fix the deployment choice

   Record the exact model version, source Region, inference geography and service tier. Keep Standard, Batch and capacity commitments in separate scenarios.

2. 2

   Count completed work and all model calls

   Estimate completed answers, model calls per answer, billed input and output, cache writes and cache reads. Include retrieval context, retries and agent loops in their actual meters.

3. 3

   Add the surrounding services

   Choose managed or self-managed retrieval, then add storage, retrieval calls, filters and any agent runtime. Give application compute, logs and networking their own lines.

4. 4

   Check a representative workload next week

   Run a small acceptance evaluation, record actual usage, and compare it with the estimate. Keep Standard until the measurements justify caching changes, batch processing or a capacity quote.

Is Amazon Bedrock Free?

Bedrock does not provide an unlimited free production token allowance. AWS advertises new-account Free Tier credits: $100 on signup and up to $100 more earned through qualifying activities. The Free plan lasts up to six months or until credits run out. Availability and credit eligibility depend on the service and account; paid-only features or a paid-plan upgrade can change what you owe. AWS Free Tier

A new user who stays within eligible credits and stops before they run out can experiment without paying out of pocket. A production application that continues consuming paid inference needs a budget after those credits. An existing startup account should not assume it qualifies for a new-account offer.

Check the expiry shown on your actual credits before relying on them. There is no Bedrock student subscription tier listed on the pricing page, and a bill should not be treated as refundable by default. Credit eligibility, billing adjustments and refunds require the applicable AWS account terms or a decision from AWS support.

Frequently Asked Questions

How much does bedrock cost?

It depends on model, Region, routing and optional services. Sonnet 5.5 global Standard is $2 input and $10 output per million tokens. The illustrative support feature above costs $1,270 monthly for 100,000 answers in the modeled Bedrock components, with application infrastructure additional.

Is there a free tier for AWS Bedrock?

Eligible new AWS accounts can use Free Tier credits for participating services. That is a limited credit allowance, not unlimited free Bedrock inference. Check model access and credit applicability in your account before relying on it.

What are EC2 pricing models?

Amazon EC2 is separately billed compute, with options including On-Demand, Savings Plans, Reserved Instances and Spot. Those are not interchangeable with Bedrock token prices or Reserved TPM capacity. Use the EC2 pricing page when your architecture includes instances.

How much does S3 cost per GB?

Amazon S3 storage prices depend on Region, storage class and volume. Requests and transfer can add charges. Use the S3 pricing page for your source documents; Bedrock Managed Knowledge Base index storage is a separate meter.

What is the pricing for 1TB of storage on Amazon S3?

There is no single all-in S3 price for that quantity across storage classes and Regions. Select the storage class and Region, then include request and transfer usage. Do not substitute the Managed Knowledge Base raw-data storage rate for an S3 estimate.

Can I get AWS S3 for free?

Eligible AWS Free Tier benefits can offset qualifying usage. Check your account's current offer and limits on the AWS Free Tier page. Free source storage would not make model inference or a Bedrock search index free.

Can I use AWS for free forever?

Only qualifying always-free service allowances can continue within their stated limits. New-account credits expire, and paid Bedrock inference is not an unlimited always-free allowance. Budget continuing usage after any applicable credits.

What are some free alternatives to Amazon S3?

For a local prototype, files you already store locally can avoid buying a cloud object-storage service. That changes the architecture: it does not supply a managed Bedrock knowledge base or hosted inference for free. For an AWS production feature, choose storage by access, durability and integration requirements.

How much does it cost to use Amazon S3?

Estimate storage, requests, retrieval where applicable and data transfer on AWS's S3 pricing page. Add that amount separately when S3 holds source documents or batch inputs and results.

Download the AI business workflow audit checklist to map one AI feature's usage, acceptance criteria and monthly budget before scaling it.