Amazon Bedrock pricing: how to cut the bill

Amazon Bedrock gives an application one API to the foundation models of Amazon, Anthropic, Mistral AI, OpenAI and other providers, and bills each call by its tokens: the input the model reads, the output it writes, and the prompt cache it reads and writes, at the rates of the model you choose. That choice usually decides most of the AWS Bedrock bill: in this article's support assistant, the model is 74.0% of it, computed, and in Paris the same request costs 216.3 times as much on the dearest model as on the cheapest. Around the tokens, a guardrail, a vector store, invocation logs and agents bill on lines of their own. This article explains what runs behind a Bedrock call, how AWS Bedrock pricing works line by line, and the ways AWS documents to reduce Bedrock costs, priced with the list prices of the Paris Region (eu-west-3), which the Bedrock pricing page does not show. Four architectures show what the levers do to whole monthly bills, and the cost section opens with an index of the levers, in the order to work through them.

What Amazon Bedrock is

In AWS's words, "Amazon Bedrock is a fully managed service that provides secure, enterprise-grade access to high-performing foundation models from leading AI companies, enabling you to build and scale generative AI applications." It "supports 100+ foundation models from industry-leading providers, including Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI, and xAI." What it bills is inference, which the key terminology defines as "The process of a foundation model generating an output (response) from a given input (prompt)", and inference is counted in tokens, each "A sequence of characters that a model can interpret or predict as a single unit of meaning."

The models share one interface. The Converse API "provides a consistent API that works with all Amazon Bedrock models that support messages. This means you can write code once and use it with different models." It follows that changing the model, the largest price lever of this article, is a parameter of the call; whether the cheaper model answers well enough is a question for the workload's own evaluation, not for the price list.

Who sells a model decides where its charges appear. Amazon's models, and those of DeepSeek, Google, Meta, MiniMax, Mistral AI, Moonshot AI, NVIDIA, Qwen, xAI and Z.AI, are not sold through AWS Marketplace (product IDs) and bill under Amazon Bedrock. The others are, Anthropic's Claude models among them, and the Claude Sonnet 4.6 model card says where they land: "Charges appear on your AWS bill and in AWS Cost Explorer under the model provider (not under Amazon Bedrock)." Bedrock subscribes the account by itself at the first call to such a model (model access). Amazon Bedrock Marketplace is a different product, which deploys models "to an endpoint managed by SageMaker AI" (Bedrock Marketplace), on instances you choose; this article does not price it. Around inference, Bedrock sells features that bill on meters of their own:

How Bedrock works under the hood

AWS documents Bedrock at the level of its endpoints, the tokens it counts, the Regions a request may run in and the tiers it sells, each a line of the bill; the features around inference are described where they are priced. How the service runs the models is not in the pages this article cites, and is left out.

One endpoint from Paris

Bedrock has two inference endpoints, bedrock-runtime and bedrock-mantle, and AWS's advice is plain: "For most new applications, use the bedrock-runtime endpoint". From Paris there is no choice to make: "The bedrock-mantle endpoint is available in a subset of those Regions" (Regional availability by endpoints), and Paris is not one of them. It follows that every call made from Paris goes through bedrock-runtime, which is also where the cost tools work: invocation logging "is only supported for calls made through the bedrock-runtime endpoint" (invocation logging), and "The bedrock-mantle endpoint doesn't support Amazon Bedrock Guardrails" (Guardrails Regions). The model providers see none of the calls: "they don't have access to Amazon Bedrock logs or to customer prompts and completions" (Data protection), and your content "is not shared with any model providers" (Bedrock FAQs).

Tokens: what one call counts

A call bills four kinds of tokens. "Amazon Bedrock charges are broken down by four token types": input, output, cache read and cache write, each with a usage type of its own, and the guide warns: "If you only sum input and output tokens, your totals will not match your bill." Output is the dear kind. In Paris an output token costs 5.0 times an input token on every Claude model, 4.0 times on Amazon Nova Micro, Lite and Pro, and 8.4 times on Amazon Nova 2 Lite, computed. A call made of a system prompt, retrieved text, a question and an answer bills like this:

Scroll the diagram sideways →

The quota does not count what the bill counts. "At the start of the request", the quota system deducts "Total input tokens + max_tokens", then adjusts to the tokens actually generated (How tokens are counted), and output tokens can weigh more on the quota than on the bill: "The burndown rate for Anthropic Claude models version 4.8 is 15x for output tokens", it is 10x for Claude Opus 5.5, Sonnet 5, Opus 5 and Fable 5.1, and 5x for the Claude models of version 4.7 and below. It follows that a large max_tokens costs nothing by itself, since the bill follows the tokens generated, but it can make requests throttle earlier, the more so on the newer models. The tokens of every call are published in Amazon CloudWatch, per model, as InputTokenCount, OutputTokenCount, CacheReadInputTokenCount and CacheWriteInputTokenCount (runtime metrics).

Where a call runs: In-Region, geographic or global

Three options decide where a request is processed. "In-Region: Your requests never leave the AWS Region you specify." A geographic profile "routes your request to a Region within a defined geography (US, Canada, EU, Asia Pacific, Japan, Australia, or India), keeping data within that geography", and a global one "to a supported commercial Region worldwide". Both are inference profiles, named after their reach: eu.anthropic.claude-sonnet-4-6 for the EU, global.anthropic.claude-sonnet-4-6 for the world. The cross-Region inference guide answers the cost question in two sentences:

There's no additional routing cost for using cross-Region inference. The price is calculated based on the Region from which you call an inference profile.

A call made from Paris and processed in Frankfurt is billed at Paris's rate. The one difference in price is the global profile's: "You save approximately 10% on both input and output token pricing compared to geographic cross-Region inference" (Global cross-Region inference), and Paris's rows give the exact share, a global row costing 9.1% less than the geographic one on every Claude model and on Amazon Nova 2 Lite, computed.

Scroll the diagram sideways →

The data goes where its profile says, with one caveat: "To the extent we store data for abuse detection, your input prompts and output results will be stored in the destination region" (Geographic cross-Region inference). Which Regions a geographic profile reaches is a property of each model: for Claude Sonnet 4.6 called from Paris, its model card lists Frankfurt, Stockholm, Milan, Spain, Ireland and Paris itself, not London. A service control policy that blocks unused Regions must allow all of a profile's destination Regions, or the requests fail.

What a caller in Paris can invoke

The availability page has one table per model and one row per Region. Its Paris rows, read on 2026-10-16, sort the models by the way Paris reaches them:

Counted from that table: seven active models run in Paris itself, all of them Mistral or embedding models; every Claude model, every Amazon Nova chat model and every OpenAI model goes through a profile; and 15 active models are reachable from Paris through the global profile only, so a workload whose data must stay in the EU cannot use them. Claude Mythos 5.1 is a gated model (its release type on the prompt caching page). The Paris prices are in On-demand tokens in Paris; the OpenAI models other than GPT-6 Astra, and Z.AI GLM 5.3, have no row in any Bedrock offer file read for this article, in Paris or in N. Virginia, and are not priced.

Service tiers: Standard, Priority, Flex and Reserved

"Amazon Bedrock offers four service tiers for model inference: Reserved, Priority, Standard, and Flex". Three are chosen per request with the service_tier parameter, Standard by default; the Reserved tier is a reservation, arranged with the AWS account team. The pricing page prices two of them against Standard, and the user guide and the service tiers page say what each changes:

Priority buys speed, "up to 25% better output tokens per second (OTPS) latency compared to Standard tier" for most of the models that offer it, in the service tiers page's words; Flex buys a discount for waiting. The Reserved tier buys availability, "The Reserved tier targets 99.5% uptime for model response", on strict terms: at least 100,000 input and 10,000 output tokens per minute (TPM), sized on input tokens and cache writes together, and "Billing continues until you delete the Reserved Tier reservation with the help of your AWS account manager." Which model offers which tier is on its model card, and every model offers Standard. From the cards of Claude Sonnet 4.6, Claude Haiku 4.5, Claude Haiku 5.5, Claude Sonnet 5.5, Amazon Nova Pro and Amazon Nova 2 Lite, with batch from the batch Regions and models page:

The Paris price list agrees with the cards: it has Priority and Flex rows for Nova Pro and Nova 2 Lite, none for Claude.

Batch inference

"With batch inference, you can submit multiple prompts and generate responses asynchronously": a job reads its prompts from Amazon S3 and writes the answers there. It gives up three things for its price: provisioned models, tool calling in Converse-format jobs, and prompt caching, which "is not supported with the batch inference API" (prompt caching). AWS gives a job no completion time, only a timeout (Create a batch inference job), and the capacity page describes batch with a "24-hour processing window". The price is the pricing page's rule, batch "at a 50% lower price compared to on-demand inference pricing", and from Paris batch runs through the EU profile for Amazon Nova 2 Lite, Nova Pro, Nova Lite and Nova Micro, and for Claude Haiku 4.5, Sonnet 4.5 and 4.6, and Opus 4.5, 4.6 and 5. Batch and Flex reach the same price by two routes:

A · Batch inference

How

A job of many prompts, read from Amazon S3, answers written back to S3

When it answers

Asynchronously: no completion time, a job timeout

Limits

No prompt caching, no tool calling in Converse-format jobs, not on provisioned models

Suits

"creating large volumes of marketing content, document classification, or data extraction"

B · Flex tier

How

An ordinary request, with service_tier set to flex

When it answers

Synchronously, after Standard requests when traffic is high

Limits

In Paris, Amazon Nova Pro and Nova 2 Lite among the models priced here; not Claude

Suits

"Real-time single API call suitable for applications that can tolerate some increase in latency but still need synchronous processing"

What differs: a file of prompts answered later, against single calls answered more slowly; the two cost the same on the models that offer both.

Prompt caching

Prompt caching reuses the start of a prompt across requests: implicitly, which "is best effort", or explicitly, with a cache checkpoint at the end of the reused prefix. "The cache has a Time To Live (TTL), which resets with each successful cache hit"; tokens read from the cache are "billed at the model's cache-read rate", and "Depending on the model, tokens written to cache can be billed at a rate that is higher than the standard input token rate." Whether a prompt can be cached at all depends on its model:

One note of the same page bears on the bill: with cross-Region inference, "At times of high demand, these optimizations may lead to increased cache writes", so a cache can hit less often through a profile than in one Region. The Paris rates are in Prompt caching in Paris.

In short: From Paris every call goes through bedrock-runtime, most of them through an inference profile, and every one is priced at Paris rates: where a request runs changes where its data goes, not its price, except for the global profile's discount.

What Amazon Bedrock is used for

AWS points each kind of work to a feature across its guides: Guardrails, prompt caching, Knowledge Bases, the service tiers page and the Agents guide:

The applications around Bedrock run on other services. In this article's patterns, the support assistant runs on AWS Lambda, whose requests and GB-seconds bill apart, and the knowledge base reads its documents from a bucket that Amazon S3 bills by the GB-month.

Bedrock or a model you host yourself

The Well-Architected Generative AI Lens frames the choice by how each is billed: "Self-hosted models on Amazon EC2 or Amazon SageMaker AI inference endpoints charge based on uptime, as well as additional costs storage and network costs" (GENCOST01-BP01), and, for one model serving small, periodic workloads, it asks to "Evaluate the cost of hosting this model on an Amazon SageMaker AI inference endpoint" and to compare it with importing the model into Bedrock (GENCOST02-BP01). Bedrock's on-demand tokens bill nothing while no request runs; the price list has no hourly row for them. The Reserved tier, Provisioned Throughput and a self-hosted model bill every hour, busy or not, as an EC2 instance does. It follows that the comparison is a question of utilization, the same as the one the Reserved tier raises inside Bedrock.

How Bedrock pricing works

What a Bedrock bill is made of

"Amazon Bedrock pricing in CUR is determined by three factors: the token type, the service tier, and whether the request was routed through cross-region inference", and the model sets the rate of each. Around the tokens, the bill has other lines, and the Paris price list does not price all of them:

The Bedrock pricing page lists no Paris Region; the four Bedrock offer files of the AWS Price List have 490 Paris products between them. No Savings Plan covers Bedrock: AWS lists "four types of Savings Plans: Compute Savings Plans, Database Savings Plans, EC2 Instance Savings Plans, and SageMaker AI" plans (Savings Plans types), and none names it; the capacity Bedrock sells is its own. Put together, the lines of a small assistant's month look like this:

On-demand tokens in Paris

Each model has its rates, and the price list writes them by the way Paris reaches the model: a row with no routing suffix prices the EU profile for a model Paris reaches through one, and the Region itself for a model that runs in Paris. The Standard tier, from the cheapest request to the dearest, a request being 2,000 input and 500 output tokens:

The newest model of a family is not the dearer one here: per token, Claude Haiku 5.5 costs 90.0% less than Claude Haiku 4.5; per request, Claude Sonnet 5.5 costs 33.3% less than Claude Sonnet 4.6, and Claude Opus 5.5 20.0% less than Claude Opus 4.6, computed. Prompts above 100,000 tokens cost more on Claude Haiku 5.5, "Prompts that exceed 100,000 tokens are billed at higher rates for input, output, and cache read and write operations" (Claude Haiku 5.5), and the Paris price list has such long-context rows for Claude Haiku 5.5 and OpenAI GPT-6 Astra among the models priced here; the table prices the shorter prompts. A step down inside one family moves a bill as much as a family change:

These are prices, not quality: a cheaper model that fails the task costs the retries and the people who fix its answers. Which model is cheapest for this job? says how to choose, and Choose the model by measured quality, not by habit is the lever.

Service tiers, batch and global routing in Paris

The tiers and the profiles change the rate of a model by fixed ratios. Input tokens, per million, for the three models Paris prices on the most tiers:

A dash is a tier the model does not offer from Paris. Output tokens follow the same ratios: batch and Flex cost 50.0% less than Standard, Priority 75.0% more, and a global row 9.1% less than the EU one, on every pair of Paris rows of the two token offer files, computed, with two exceptions: three batch rows of a preview model, Amazon Nova 2 Omni, rounded a little below half, and TwelveLabs Pegasus v1.2, which costs the same through both profiles.

Prompt caching in Paris

A cached prefix bills twice: once when it is written, at a write rate, then at a read rate each time a request finds it. Per million tokens:

On Claude, a write with the 5-minute TTL costs 1.25 times the input rate and a read 0.1 times, computed; the 1-hour TTL writes at 2 times, $6.60 per million on Claude Sonnet 4.6 and $2.20 on Claude Haiku 4.5. It follows that a prefix pays for its 5-minute write at its first read, since one read saves more than the write cost extra, and for a 1-hour write at its second read; a prefix written and never read costs more than no cache at all. On Amazon Nova the write costs nothing and a read 0.25 times the input rate, which the pricing page says for Nova 2 Lite: "Cache read input tokens will be 75% less than on-demand input token price". Pattern 3 prices a cache on a whole month.

The Reserved tier in Paris

The Reserved tier prices capacity, not tokens: an amount per 1,000 tokens per minute of input and of output, for every hour of the term.

The prices hide a simple rule. 1,000 tokens a minute for an hour is 60,000 tokens, and 60,000 Standard tokens of Claude Sonnet 4.6 through the EU profile cost exactly the 1-month price of 1,000 TPM for an hour, input and output alike; the same holds for the global profile, for Claude Haiku 4.5 and for the other Claude models with Reserved rows, and a 3-month term costs nine tenths of the 1-month one, computed. It follows that a 1-month reservation costs what Standard tokens cost at full use of the capacity, and never less, and that a 3-month reservation costs less only above this utilization:

Below it, the Reserved tier buys availability and a capacity guarantee, not a discount. Pattern 4 prices a reservation used at 95%, and Commit only to measured, steady capacity the smallest one AWS sells.

Provisioned Throughput in Paris

Provisioned Throughput holds capacity in model units, billed by the hour: "You're billed hourly for a Provisioned Throughput that you purchase", a 1- or 6-month term lowers the rate, and what a unit delivers comes from the account team, not from a page. "If you customized a model, you must purchase Provisioned Throughput to be able to use it", and "Inference profiles currently don't support Provisioned Throughput" (cross-Region inference), so no model Paris reaches through a profile can use it. In Paris the supported models page lists Titan Multimodal Embeddings G1 and the two Cohere Embed models, and the price list has rows for the first only among current models; the purchase page also names "two types of Provisioned Throughput - by Tokens and by Model Units", the first described in no page read for this article.

A 1-month term costs 9.9% less per hour than no commitment, and a 6-month term 28.0% less, computed. With no published throughput per unit, no break-even against the model's on-demand $1.00 per million tokens can be computed: one unit on a 1-month term costs as much as $7,832.90 of on-demand tokens a month, and whether it processes that many is a question for the account team.

Knowledge bases and vector stores in Paris

"With Amazon Bedrock Knowledge Bases, you can integrate proprietary information into your generative-AI applications": a sync splits your documents into chunks that an embedding model turns into vectors, and at query time the question is embedded too and the closest chunks join the prompt (How knowledge bases work); "If you modify the data in the data source, you must resync the changes" (Build a knowledge base). There are two kinds. A managed knowledge base bills its own index storage and retrievals, parsing and embeddings included, and is offered in eight Regions (managed knowledge base Regions), Frankfurt, Ireland and London among them, not Paris. A knowledge base with a vector store keeps its vectors in a store you create and pay for, Amazon OpenSearch Serverless, Amazon S3 Vectors, Amazon Aurora or four others (vector store prerequisites), and the pricing page prices nothing for it but structured-data queries and rerank models. It follows that a knowledge base built in Paris bills the embedding model's tokens, the vector store's own service and the answering model's tokens, and nothing for the knowledge base itself:

Scroll the diagram sideways →

The two stores bill on opposite principles. Amazon OpenSearch Service runs OpenSearch Serverless, which bills compute by the hour, in OpenSearch Compute Units (OCUs), each "a combination of 6 GiB of memory and corresponding virtual CPU (vCPU)" (capacity limits), with a floor for classic collections: "You will be billed at least for a minimum of 2 OCUs … for the first collection in an account" (OpenSearch pricing). NextGen collections have none, and whether Paris offers them, or a knowledge base's quick-create flow makes one, the pages read do not say. S3 Vectors bills storage and each query instead, and Paris has its endpoint (S3 Vectors Regions). In Paris a knowledge base can use three embedding models, Titan Text Embeddings V2, Cohere Embed English and Cohere Embed Multilingual (supported models); the first offers 256, 512 or 1,024 dimensions.

The two stores part ways at the floor. A classic OpenSearch Serverless collection at its minimum bills every hour of the month:

An S3 Vectors query bills its request and the data it scans, "The data processed charge is based on the size of your vector index" (Amazon S3 pricing): on pattern 1's index of 100,000 vectors, S3 Vectors reaches the OpenSearch minimum at about 90 million queries a month, computed. Embedding 50 million tokens of documents costs $1.50 with Titan Text Embeddings V2 and $6.00 with Cohere Embed v4 through the EU profile.

Guardrails in Paris

Amazon Bedrock Guardrails "provides configurable safeguards to help you build safe generative AI applications". Only the policies a guardrail has bill, each on every text it evaluates, and a blocked prompt costs the guardrail but not the model: "If a guardrail blocks the input prompt, you're charged for the guardrail evaluation. There are no charges for foundation model inference calls" (How Guardrails works). The unit is text: "A text unit can contain up to 1000 characters", and "if a text input contains 5600 characters, it will be charged for 6 text units" (Amazon Bedrock pricing). Per policy:

Word filters and regular expressions cost nothing; the image content filter, which the pricing page prices "per image processed", has no Paris row. The Guardrails Standard and Classic tiers are safeguard tiers, not the service tiers above: the Classic one supports English, French and Spanish (safeguard tiers), and "The pricing is the same for both standard tier and classic tier". Guardrail what needs it, and only that prices a guardrail by what it reads.

Amazon Bedrock AgentCore in Paris

Bedrock's first agent service is closing: "Amazon Bedrock Agents (launched November 2023) is now Amazon Bedrock Agents Classic and will no longer be open to new customers starting on July 30, 2026" (Agents Classic maintenance mode), with no migration deadline and no charge of its own. AWS recommends Amazon Bedrock AgentCore instead, which its pricing page bills by use, "with no upfront commitments or minimum fees": the Runtime's CPU and memory by the second, gateway calls, memory records and evaluations. Identity used through the Runtime or the Gateway costs nothing more, and Observability bills as Amazon CloudWatch. Paris offers all of it but the Runtime's instances, the AWS Agent Registry and the Web Search tool (AgentCore Regions):

The pricing page adds that "128MB minimum memory billing applies for memory", and that for built-in evaluators "Model usage costs are included". The Paris price list has no row for the Runtime's v2 rates, which the pricing page describes, or for Policy, and its short-term memory row is priced per event at zero where the page bills per GB ingested: this article prices none of them. An agent's model calls bill as the tokens of On-demand tokens in Paris.

Prompt routing, prompt optimization and evaluations

Evaluations bill the models they run: "If you use RAG evaluation or LLM-as-a-judge in Model Evaluation, the tokens that the judge model uses are charged based on the on-demand standard tier prices", and human evaluation adds a charge per completed human task, which the pricing page gives and the price list does not. Model customization has no Paris row: Distill or customize a smaller model, where it is offered gives N. Virginia's for scale. Custom model import and Bedrock Data Automation have Frankfurt rows and no Paris one; this article does not price them.

The costs around Bedrock

Model invocation logs, off by default, go to Amazon CloudWatch Logs or Amazon S3 (invocation logging), and they hold whole prompts and answers, so their volume follows the tokens. CloudWatch's list of vended log sources names Bedrock knowledge base logging, agent logging and the AgentCore components, not model invocation logs: it follows that those bill at the custom-log rate above. AgentCore's telemetry goes to CloudWatch too: "Standard CloudWatch rates apply if Observability is enabled". The S3 article prices the requests that batch jobs and data sources make.

Paris and N. Virginia

The Bedrock pricing page shows US and other Regions, not Paris, and the guides that copy it price N. Virginia. For Claude that is the Paris price; for Amazon's own models it is not:

It follows that an estimate made from N. Virginia's rates holds for a Claude workload in Paris and understates an Amazon Nova one by about a third.

In short: In Paris the model sets the price of every token, the tiers move it by fixed ratios, and the Reserved tier is a purchase of availability rather than a discount unless it is used above 90.0%, computed.

Choosing a model and a capacity mode

The lens puts the first fact plainly: "Foundation model costs vary greatly across the various foundation model providers, model families and sizes, and model hosting paradigms". Four choices decide a Bedrock bill, each with its answer in Paris.

Which model is cheapest for this job?

The cheapest request is on Amazon Nova Micro and the dearest on Claude Fable 5.1 (On-demand tokens in Paris), but the cheapest model for a job is the cheapest one that passes the job's own evaluation. Among the models a support or chat workload compares, Claude Sonnet 4.6 costs 3.0 times Claude Haiku 4.5 per request, computed, and Claude Haiku 5.5 costs less again; Mistral 7B Instruct, which runs in Paris, costs about what Claude Haiku 5.5 costs through the EU profile. Choose the model by measured quality, not by habit says how to test.

Geographic or global cross-Region inference?

Geographic when the data must stay in the EU, global otherwise: "Choose Geographic cross-Region inference when you have data residency requirements and need to ensure data processing remains within specific geographic boundaries" (cross-Region guide). The saving is the same share on every model with both rows in Paris:

Amazon Nova Pro, Nova Lite and Nova Micro have no global profile from Paris, and the models Paris reaches through the global profile alone are out of reach for data that must stay in the EU.

Standard, Flex, batch or Priority?

By how long the answer can wait: Standard, or Priority where the model offers it, for a user waiting on the reply; Flex for a call that must stay synchronous but can be slow; batch for a file of answers that can wait (Batch inference). In Paris, Flex and Priority exist for Amazon Nova Pro and Nova 2 Lite among the models priced here, not for Claude.

On demand, the Reserved tier or Provisioned Throughput?

On demand, unless the traffic is steady, high and near the capacity it would reserve (The Reserved tier in Paris); below the break-even, what a reservation buys is availability. The two ways to hold capacity:

A · Reserved tier

Unit

1,000 tokens per minute, input and output apart, per hour

Terms

1 or 3 months, at least 100,000 input and 10,000 output TPM

From Paris

Claude models, through the EU or global profile: $0.198 per 1,000 input TPM-hour on Claude Sonnet 4.6

Above the capacity

Overflow to the Standard tier, billed per token

B · Provisioned Throughput

Unit

A model unit per hour, its throughput not published

Terms

No commitment, 1 month or 6 months

From Paris

Titan Multimodal Embeddings G1 among current models: $10.73 per model unit-hour on a 1-month term

Above the capacity

Not described in the pages read; no inference profile can use it

What differs: a reservation counted in tokens per minute for the models Paris reaches through a profile, against model units for a few models served in Paris itself.

Provisioned Throughput serves a few embedding models in Paris; the models you customize are served this way in the Regions that offer customization.

In short: Choose the model by evaluation, the profile by where the data may go, the tier by how long the answer can wait, and capacity only for traffic that would keep it busy.

Architecture patterns around Bedrock and what they cost

Four illustrative configurations in Paris, each priced before and after the levers of the next section, line by line from the AWS Price List: a support assistant on a knowledge base, whose model, guardrail and vector store all move; a nightly document job, priced on demand, in batch and on Flex; a chat product whose long system prompt repeats on every request; and steady traffic, on the Standard tier against a reservation. None is a client case; every amount is computed, and each card lists its assumptions and what it leaves out. The Lambda patterns price the application side the same way.

Pattern 1 of 4 · Illustrative configuration

A support assistant on a knowledge base

A customer-support assistant in Paris answers 300,000 questions a month from a knowledge base of 20,000 help articles, embedded with Titan Text Embeddings V2 and stored in a vector store; a guardrail checks the request and Claude answers through the EU inference profile.

Scroll the diagram sideways →

Before the levers, the model is 74.0% of the bill, the guardrail 22.9% and the vector store 3.1%, computed; the embeddings are a rounding error. The model is the largest step; in another order the steps change size, the total does not. A managed knowledge base, priced in Frankfurt since Paris does not offer one, would cost $310.00 a month for the same articles and retrievals, against $2.02 for the S3 Vectors store and the embeddings in Paris. The first embedding of the 50 million tokens of articles costs $1.50 once, and the first upload to S3 Vectors $0.12.

Before · Claude Sonnet 4.6, no cache, OpenSearch Serverless, guardrail on the whole prompt$13,370.87

With the levers · Claude Haiku 4.5, cached prefix, S3 Vectors, guardrail on the question$2,124.97

• Compute$9,900.00 → $1,852.95, too small to draw
• Storage$0.01 → $0.04, too small to draw
• Other$3,470.86 → $271.98, too small to draw

Before · Claude Sonnet 4.6, no cache, OpenSearch Serverless, guardrail on the whole prompt: compute $9,900.00, storage $0.01, other $3,470.86, total $13,370.87 a month.

With the levers · Claude Haiku 4.5, cached prefix, S3 Vectors, guardrail on the question: compute $1,852.95, storage $0.04, other $271.98, total $2,124.97 a month.

With the levers: $11,245.90 a month less (84.1% less), computed.

A support assistant on a knowledge base, one lever at a time: the bill a month, computed

• Before and along the way
• With every lever

Step		A month, computed	Change
Claude Sonnet 4.6, no cache, OpenSearch Serverless, guardrail on the whole promptBefore the levers		$13,370.87	—
Guardrail on the questionLever: Guardrail what needs it, and only that		$10,580.87	−$2,790.00
Claude Haiku 4.5Lever: Choose the model by measured quality, not by habit		$3,980.87	−$6,600.00
Prompt cache, 5-minute TTLLever: Cache the prefix that repeats		$2,533.82	−$1,447.05
S3 VectorsLever: Pick the vector store for the query volume		$2,124.97	−$408.85

USD, before tax. AWS Price List, eu-west-3, offer files published October 12, 2026, October 14, 2026, October 15, 2026 and October 16, 2026, fetched October 16, 2026.

Levers

• Guardrail what needs it, and only that
• Choose the model by measured quality, not by habit
• Cache the prefix that repeats
• Pick the vector store for the query volume

The bills line by line, the assumptions and the sources

A support assistant on a knowledge base, before: Claude Sonnet 4.6, no cache, OpenSearch Serverless, guardrail on the whole prompt Illustrative

The EU profile; an OpenSearch Serverless Classic collection at its minimum; the guardrail evaluates the whole prompt and the answer.

Line	A month
Compute	$9,900.00
Claude Sonnet 4.6 input, no cache2,400 M tokens × $3.30 /M tok	$7,920.00
Claude Sonnet 4.6 output120 M tokens × $16.50 /M tok	$1,980.00
Storage
OpenSearch Serverless, managed storage0.5 GB × $0.024 /GB-mo	$0.01
Other	$3,470.86
Embeddings, questions, Titan Text Embeddings V215 M tokens × $0.03 /M tok	$0.45
Embeddings, help-article resync (10% of the index a month)5 M tokens × $0.03 /M tok	$0.15
OpenSearch Serverless, indexing OCU (Classic minimum)730 OCU-h × $0.281 /OCU-h	$205.13
OpenSearch Serverless, search OCU (Classic minimum)730 OCU-h × $0.281 /OCU-h	$205.13
Guardrail, content filters (the whole prompt and the answer)10,200,000 text units × $0.15 /1K TU	$1,530.00
Guardrail, denied topics (the whole prompt and the answer)10,200,000 text units × $0.15 /1K TU	$1,530.00
A month, computed	$13,370.87

A support assistant on a knowledge base, with the levers: Claude Haiku 4.5, cached prefix, S3 Vectors, guardrail on the question Illustrative

The EU profile; the system prompt cached with a 5-minute TTL; an S3 vector index; the guardrail evaluates only the question and the answer.

Line	A month
Compute	$1,852.95
Claude Haiku 4.5 input, uncached (retrieved text and the question)900 M tokens × $1.10 /M tok	$990.00
Claude Haiku 4.5 prompt cache write, 5-minute TTL30 M tokens × $1.375 /M tok	$41.25
Claude Haiku 4.5 prompt cache read1,470 M tokens × $0.11 /M tok	$161.70
Claude Haiku 4.5 output120 M tokens × $5.50 /M tok	$660.00
Storage
S3 Vectors, storage0.57 GB × $0.063 /GB-mo	$0.04
Other	$271.98
Embeddings, questions, Titan Text Embeddings V215 M tokens × $0.03 /M tok	$0.45
Embeddings, help-article resync (10% of the index a month)5 M tokens × $0.03 /M tok	$0.15
S3 Vectors, PUT (resynced vectors)0.06 GB × $0.21 /GB	$0.01
S3 Vectors, queries300,000 requests × $0.0026 /1K req	$0.78
S3 Vectors, data processed by queries (first 100,000 vectors of the index)143,051.15 GB × $0.000004092 /GB	$0.59
Guardrail, content filters (the question and the answer)900,000 text units × $0.15 /1K TU	$135.00
Guardrail, denied topics (the question and the answer)900,000 text units × $0.15 /1K TU	$135.00
A month, computed	$2,124.97

Claude Sonnet 4.6 with the cache, S3 Vectors and the guardrail on the question: $5,830.87 a month, 56.4% less than before, computed.

Assumptions

• Illustrative configuration in Paris (eu-west-3), not a client case.
• 300,000 questions a month (about 10,000 a day). The knowledge base indexes 20,000 help articles in an S3 bucket, split into 100,000 chunks of about 500 tokens (50,000,000 tokens); a tenth of the chunks is re-embedded each month.
• Each question: the knowledge base embeds it (50 tokens) and returns 5 chunks; the model receives a 5,000-token static prefix (instructions, policies, tool definitions), 2,500 tokens of chunks and 500 tokens of question and history, and answers in 400 tokens. All assumptions.
• Before: an OpenSearch Serverless Classic collection at its minimum of 2 OCUs (an assumption that the minimum serves this index); the guardrail evaluates the whole prompt and the answer.
• After: the 5,000-token prefix is cached with the 5-minute TTL (above Claude Haiku 4.5's 4,096-token minimum), written on 2% of requests (cold cache: nights included) and read on 98%; an S3 vector index of 100,000 vectors, 6 KB stored and 5 KB processed per vector per query (4 KB of 1,024-dimension float32 vector data, the chunk's metadata and key: an assumption); the guardrail evaluates only the question and the answer. Claude Haiku 4.5 passing the team's evaluation is an assumption to test.

Not counted

• The application (an AWS Lambda function, the pattern assumes), the S3 bucket of articles, CloudWatch, data transfer, the free tiers.

Sources

• Bedrock User Guide, "Regional availability by models"
• Bedrock User Guide, model card "Claude Sonnet 4.6"
• Bedrock User Guide, model card "Claude Haiku 4.5"
• Bedrock User Guide, "Prompt caching for faster model inference"
• Bedrock User Guide, "Prerequisites for using a vector store you created for a knowledge base"
• Amazon OpenSearch Service pricing page
• Amazon S3 pricing page
• Bedrock User Guide, "Supported models and Regions for Amazon Bedrock knowledge bases"
• Bedrock User Guide, "How Amazon Bedrock Guardrails works"
• Bedrock User Guide, "Apply tags to user input to filter content"
• Amazon Bedrock pricing page
• Well-Architected Generative AI Lens, "GENCOST03-BP04"
• Well-Architected Generative AI Lens, "GENCOST01-BP01"

USD before tax, 730 h a month. AWS Price List, eu-west-3, offer files published October 12, 2026, October 14, 2026, October 15, 2026 and October 16, 2026, fetched October 16, 2026. Illustrative configuration, not a client case.

Pattern 2 of 4 · Illustrative configuration

A nightly document job: on demand, batch or Flex

Every night 20,000 documents (contracts, tickets) are summarized and classified on Amazon Nova Pro through the EU profile: 600,000 documents a month, 3,000 input tokens and 300 output tokens each. Results are needed the next morning.

Scroll the diagram sideways →

The Flex tier gives the same bill without the job, for calls that must stay synchronous. A smaller model in batch, if it passes the evaluation, moves the bill further than the tier: the card's details price Claude Haiku 4.5, Amazon Nova 2 Lite and Amazon Nova Micro batch jobs, and Priority, the other direction.

Before · Nova Pro on demand, Standard tier$2,973.60

With the levers · Nova Pro batch jobs$1,486.80

• Compute$2,973.60 → $1,486.80, too small to draw

Before · Nova Pro on demand, Standard tier: compute $2,973.60, total $2,973.60 a month.

With the levers · Nova Pro batch jobs: compute $1,486.80, total $1,486.80 a month.

With the levers: $1,486.80 a month less (50.0% less), computed.

Levers

• Pick the tier by how long the answer can wait

The bills line by line, the assumptions and the sources

A nightly document job: on demand, batch or Flex, before: Nova Pro on demand, Standard tier Illustrative

Synchronous calls during the night, Standard tier.

Line	A month
Compute	$2,973.60
Amazon Nova Pro input, 1800 million tokens1,800 M tokens × $1.18 /M tok	$2,124.00
Amazon Nova Pro output, 180 million tokens180 M tokens × $4.72 /M tok	$849.60
A month, computed	$2,973.60

A nightly document job: on demand, batch or Flex, with the levers: Nova Pro batch jobs Illustrative

One batch inference job a night, input and output JSONL in S3.

Line	A month
Compute	$1,486.80
Amazon Nova Pro input, 1800 million tokens1,800 M tokens × $0.59 /M tok	$1,062.00
Amazon Nova Pro output, 180 million tokens180 M tokens × $2.36 /M tok	$424.80
A month, computed	$1,486.80

Nova Pro on the Flex tier: $1,486.80 a month, 50.0% less than before, computed.

Claude Haiku 4.5 batch jobs: $1,485.00 a month, 50.1% less than before, computed.

Amazon Nova 2 Lite batch jobs: $800.91 a month, 73.1% less than before, computed.

Amazon Nova Micro batch jobs: $65.52 a month, 97.8% less than before, computed.

Nova Pro on the Priority tier: $5,203.80 a month, 75.0% more than before, computed.

Assumptions

• Illustrative configuration in Paris (eu-west-3), not a client case.
• 600,000 documents a month (20,000 a night), 3,000 input tokens and 300 output tokens each (1,800 million input and 180 million output tokens a month). All assumptions.
• Before: synchronous calls during the night, Standard tier. After: one batch inference job a night, input and output JSONL in S3.
• A job finishing within the night (AWS states no completion time for a job, only a 24-hour processing window and a job timeout). No tool calling in the job (not supported for Converse-format batch jobs).

Not counted

• The S3 storage and requests of the JSONL files (small, billed by S3), the scheduler and the function that starts the job, the free tiers.

Sources

• Bedrock User Guide, "Process multiple prompts with batch inference"
• Bedrock User Guide, "Supported Regions and models for batch inference"
• Bedrock User Guide, "Create a batch inference job"
• Bedrock User Guide, "Service tiers for optimizing performance and cost"
• Amazon Bedrock service tiers page
• Amazon Bedrock pricing page
• Bedrock User Guide, model card "Nova Pro"

USD before tax, 730 h a month. AWS Price List, eu-west-3, offer files published October 12, 2026 and October 14, 2026, fetched October 16, 2026. Illustrative configuration, not a client case.

Pattern 3 of 4 · Illustrative configuration

A chat product with a long shared system prompt

A chat product sends 500,000 requests a month to Claude Sonnet 4.6 through the EU profile. Every request starts with the same 10,000-token system prompt, then 1,000 tokens of conversation; answers are 300 tokens. Traffic is steady in the day and sparse at night.

Scroll the diagram sideways →

The chart prices the cache by the share of requests that find it cold and write the prefix again. The cache costs as much as no cache only when 78.3% of requests write it, computed, and a cache written on every request and never read costs more than none. The 1-hour TTL, written less often, and the global profile, in the card's details, cut the bill further.

Before · No prompt cache$20,625.00

With the levers · Prompt cache, 5-minute TTL$6,154.50

• Compute$20,625.00 → $6,154.50, too small to draw

Before · No prompt cache: compute $20,625.00, total $20,625.00 a month.

With the levers · Prompt cache, 5-minute TTL: compute $6,154.50, total $6,154.50 a month.

With the levers: $14,470.50 a month less (70.2% less), computed.

A chat product's month by the share of requests that write the cached prefix, against no cache, computed

• No prompt cache
• With the cache

Share that writes		No prompt cache	With the cache	Change, computed
0% write the prefix		$20,625.00	$5,775.00	72.0% less
2% write the prefixThis pattern		$20,625.00	$6,154.50	70.2% less
10% write the prefix		$20,625.00	$7,672.50	62.8% less
25% write the prefix		$20,625.00	$10,518.75	49.0% less
50% write the prefix		$20,625.00	$15,262.50	26.0% less
100% write the prefix		$20,625.00	$24,750.00	20.0% more

Caching costs the same as no cache at 78.3% of requests writing the prefix, computed.

USD, before tax. AWS Price List, eu-west-3, offer file published October 12, 2026, fetched October 16, 2026.

Levers

• Cache the prefix that repeats

The bills line by line, the assumptions and the sources

A chat product with a long shared system prompt, before: No prompt cache Illustrative

Every request pays the 10,000-token system prompt at the input rate.

Line	A month
Compute	$20,625.00
Input at the input rate (system prompt and conversation)5,500 M tokens × $3.30 /M tok	$18,150.00
Answer150 M tokens × $16.50 /M tok	$2,475.00
A month, computed	$20,625.00

A chat product with a long shared system prompt, with the levers: Prompt cache, 5-minute TTL Illustrative

A cache checkpoint after the system prompt; the prefix is written on 2% of requests and read on 98%.

Line	A month
Compute	$6,154.50
Conversation turns, billed as input500 M tokens × $3.30 /M tok	$1,650.00
Prompt cache write, 2% of requests100 M tokens × $4.125 /M tok	$412.50
Prompt cache read, 98% of requests4,900 M tokens × $0.33 /M tok	$1,617.00
Answer150 M tokens × $16.50 /M tok	$2,475.00
A month, computed	$6,154.50

1-hour TTL, 0.2% of requests write: $5,837.70 a month, 71.7% less than before, computed.

Global profile with the 5-minute cache: $5,595.00 a month, 72.9% less than before, computed.

A cache written on every request and never read: $24,750.00 a month, 20.0% more than before, computed.

Assumptions

• Illustrative configuration in Paris (eu-west-3), not a client case.
• 500,000 requests a month, each starting with the same 10,000-token system prompt (instructions, product documentation, tool definitions), then 1,000 tokens of conversation; answers are 300 tokens. All assumptions.
• Before: every request pays the system prompt at the input rate. After: a cache checkpoint after the system prompt (above the 1,024-token minimum), the 5-minute TTL; the prefix is written on 2% of requests (cold cache: the first request of the day, gaps over 5 minutes) and read on 98%.

Not counted

• The application, logs, the free tiers.

Sources

• Bedrock User Guide, "Prompt caching for faster model inference"
• Bedrock User Guide, model card "Claude Sonnet 4.6"
• Well-Architected Generative AI Lens, "GENCOST03-BP03"

USD before tax, 730 h a month. AWS Price List, eu-west-3, offer file published October 12, 2026, fetched October 16, 2026. Illustrative configuration, not a client case.

Pattern 4 of 4 · Illustrative configuration

Steady traffic: Standard tokens against the Reserved tier

An application sends Claude Sonnet 4.6 (EU profile) a steady flow around the clock: on average 190,000 input and 19,000 output tokens a minute, 95% of a 200,000 / 20,000 TPM reservation.

Scroll the diagram sideways →

At the pattern's 95% average use, the 3-month reservation costs a little less than the same tokens on the Standard tier; at 90.0%, the two cost the same, and below it the reservation costs more, computed. The 1-month term costs more at any use below full, and the global profile without any commitment, in the card's details, saves more than the reservation does through the EU profile.

Before · Standard tier, EU profile$41,193.90

With the levers · Reserved tier, 3 months, EU profile$39,025.80

• Compute$41,193.90 → $39,025.80, too small to draw

Before · Standard tier, EU profile: compute $41,193.90, total $41,193.90 a month.

With the levers · Reserved tier, 3 months, EU profile: compute $39,025.80, total $39,025.80 a month.

With the levers: $2,168.10 a month less (5.3% less), computed.

Steady traffic's month on the Standard tier by average utilization of the reservation, against a 3-month Reserved tier reservation, computed

• Standard tier
• Reserved tier, 3 months

Average utilization		Standard tier	Reserved tier, 3 months	Change, computed
50% average use		$21,681.00	$39,025.80	80.0% more
55% average use		$23,849.10	$39,025.80	63.6% more
60% average use		$26,017.20	$39,025.80	50.0% more
65% average use		$28,185.30	$39,025.80	38.5% more
70% average use		$30,353.40	$39,025.80	28.6% more
75% average use		$32,521.50	$39,025.80	20.0% more
80% average use		$34,689.60	$39,025.80	12.5% more
85% average use		$36,857.70	$39,025.80	5.9% more
90% average useBreak-even		$39,025.80	$39,025.80	the same
95% average useThis pattern		$41,193.90	$39,025.80	5.3% less
100% average use		$43,362.00	$39,025.80	10.0% less

The reservation's price does not change with use; the Standard bill is for the same traffic at each utilization.

USD, before tax. AWS Price List, eu-west-3, offer file published October 12, 2026, fetched October 16, 2026.

Levers

• Commit only to measured, steady capacity

The bills line by line, the assumptions and the sources

Steady traffic: Standard tokens against the Reserved tier, before: Standard tier, EU profile Illustrative

Tokens on demand, no commitment.

Line	A month
Compute	$41,193.90
Standard tier input8,322 M tokens × $3.30 /M tok	$27,462.60
Standard tier output832.2 M tokens × $16.50 /M tok	$13,731.30
A month, computed	$41,193.90

Steady traffic: Standard tokens against the Reserved tier, with the levers: Reserved tier, 3 months, EU profile Illustrative

200,000 input and 20,000 output TPM reserved, billed every hour of the term.

Line	A month
Compute	$39,025.80
Reserved tier input, 3 months200 1K TPM × 730 h × $0.1782 /1K TPM-h	$26,017.20
Reserved tier output, 3 months20 1K TPM × 730 h × $0.891 /1K TPM-h	$13,008.60
A month, computed	$39,025.80

Reserved tier, 1 month, EU profile: $43,362.00 a month, 5.3% more than before, computed.

Standard tier, global profile: $37,449.00 a month, 9.1% less than before, computed.

Reserved tier, 3 months, global profile: $35,478.00 a month, 13.9% less than before, computed.

Assumptions

• Illustrative configuration in Paris (eu-west-3), not a client case.
• 190,000 input and 19,000 output tokens a minute on average, 95% of a 200,000 / 20,000 TPM reservation (8,322 million input tokens a month at that average). All assumptions.
• No prompt cache (cache writes would count against the reservation, and whether a cache read served while a reservation is in force bills at the Reserved price or at Standard's cache-read rate is not stated in the pages read: an open question). The reservation is granted by the AWS account team.

Not counted

• The application, the free tiers.

Sources

• Bedrock User Guide, "Capacity and Performance"
• Bedrock User Guide, model card "Claude Sonnet 4.6"
• Bedrock User Guide, "Route model inference requests across AWS Regions with cross-Region inference"
• Bedrock User Guide, "Global cross-Region inference"
• Well-Architected Generative AI Lens, "GENCOST02-BP01"
• Bedrock User Guide, "Increase model invocation capacity with Provisioned Throughput in Amazon Bedrock"
• Bedrock User Guide, "Supported Region and models for Provisioned Throughput"

USD before tax, 730 h a month. AWS Price List, eu-west-3, offer file published October 12, 2026, fetched October 16, 2026. Illustrative configuration, not a client case.

How to reduce Bedrock costs

Each lever below comes from AWS documentation; their order is this article's. A reservation is a commitment billed every hour of its term, so it comes last, once the tokens it would cover are as few as they will get: see the bill first, trim what is billed but serves nothing, reshape the model, the prompt and the tier, route traffic through the cheapest profile the data allows, and commit to what remains. The one that moves the bill most is the model: Choose the model by measured quality, not by habit is pattern 1's largest step. The levers, in order:

1. Lever 01,SeeSee the bill by model, token type and applicationAlways, and first
2. Lever 02,TrimCap the answer, trim the promptEvery prompt template
3. Lever 03,TrimGuardrail what needs it, and only thatGuardrails on long, trusted prompts
4. Lever 04,TrimStop agents that run longAgents and multi-step workflows
5. Lever 05,TrimLog invocations on purposeHigh-volume logged applications
6. Lever 06,ReshapeChoose the model by measured quality, not by habitEvery workload, and at each model release
7. Lever 07,ReshapeCache the prefix that repeatsA static prefix reused within its TTL
8. Lever 08,ReshapePick the tier by how long the answer can waitWork that can wait, and work that cannot
9. Lever 09,ReshapePick the vector store for the query volumeKnowledge bases with modest indexes
10. Lever 10,ReshapeDistill or customize a smaller model, where it is offeredA narrow, high-volume task
11. Lever 11,RouteRoute globally when the data may leave the EUNo data-residency constraint
12. Lever 12,RouteRoute between the models of a familyMixed traffic, mostly easy requests
13. Lever 13,CommitCommit only to measured, steady capacitySteady traffic near the reservation
14. Lever 14,ReviewReview on a cadence, and at every model changeQuarterly, and at every model change

In short: Commit last: see the bill by model and application, trim the answers, the guardrail's reading and the logs, choose the model by evaluation, cache and batch what repeats or waits, route globally where the data may go, and only then reserve capacity for the traffic that remains.

FAQ

How is Amazon Bedrock billed?

Per token: input, output, cache write and cache read, at the rates of the model, the service tier and the routing of each call, priced at the rates of the Region the call is made from. Guardrails, knowledge bases, AgentCore and reserved capacity add their own lines, and Claude bills under its provider through AWS Marketplace (What a Bedrock bill is made of).

Which models can I use from Paris, and at what price?

Mistral and embedding models run in Paris itself; Claude, Amazon Nova and most others are reached through the EU or the global profile, some through the global profile only, all at Paris rates (What a caller in Paris can invoke). Input tokens run from $0.052 per million on Amazon Nova Micro to $10.00 on Claude Fable 5.1, and one request costs 216.3 times as much on the second as on the first, computed (On-demand tokens in Paris).

Is Amazon Bedrock free?

No. The price list has no monthly allowance of free tokens; new AWS accounts receive Free Tier credits, which "expire 6 months from the date you create your AWS account" (Free Tier FAQs), and this article does not deduct them. No Savings Plan covers Bedrock either (What a Bedrock bill is made of).

Does prompt caching save money?

When the prefix is read again within its TTL: on Claude a read costs 0.1 times the input rate and a 5-minute write 1.25 times, so the first read repays the write, computed. A cache written and never read costs more than none (Prompt caching in Paris).

Is Provisioned Throughput or the Reserved tier worth it?

For availability, possibly; for money, rarely. In Paris a 1-month reservation never costs less than Standard tokens and a 3-month one only above 90.0% average use, computed; Provisioned Throughput serves a few models there and its unit's throughput is not published (The Reserved tier in Paris).