Amazon Bedrock pricing: tokens, tiers, caching and batch

Amazon Bedrock prices most models by the token: one price per million input tokens and one per million output tokens, per model and per Region. Three choices then move that price: the inference profile (Global, or a Geo profile that keeps processing in the US, EU or another geography for 10% more), the service tier (Flex at half price, Priority at 75% more, Reserved capacity by the month) and prompt caching, which bills the repeated part of a prompt at a small fraction of the input price.

There is no fee for Bedrock itself. You pay for what the models read and write, plus any extras you switch on: GuardrailGuardrailChecks placed around a model that stop or change what goes in and what comes out, such as a filter on user messages or an approval step before an agent acts. checks, knowledge bases, agent tools. This article covers the model bill, which is most of it.

The prices that matter today

For the newest Claude models on the Global profile, called from US East (N. Virginia), on 22 October 2026, in dollars per million tokens:

Haiku 5.5 has a second, long-context price: a request over 100,000 input tokens costs $0.50 in and $2.50 out, 5 times the normal rate. These are the same as Anthropic's own list prices, which is why Claude on Bedrock or the Claude API is a choice about features and billing, not price. Amazon's own Nova 2 Lite costs $0.30 in and $2.50 out. Each model's page in our price tables keeps the current figures.

Global or Geo: the 10% for keeping data in one place

Every Claude call goes through an inference profile. A Global profile (global.anthropic.claude-sonnet-5-5) lets Bedrock run the request in any commercial AWS Region. A Geo profile (us., eu., au.) keeps processing inside that geography and costs 10% more: Sonnet 5.5 becomes $2.20 and $11. There is no separate fee for routing across Regions; the price is set by the Region you call from, and CloudTrail records where each request actually ran.

So the 10% is the price of a data residency rule. If your contracts or regulators require EU processing, pay it. If nothing requires it, Global is cheaper and has more capacity behind it.

Service tiers: Flex, Standard, Priority and Reserved

You choose a tier per request with the service_tier parameter:

• Standard (default, or no parameter) is the normal price.
• Flex is half the Standard price for work that can wait: evaluations, summaries, overnight jobs.
• Priority costs 75% more and is served first, for customer-facing calls that cannot be slow.
• Reserved buys input and output tokens per minute for 1 or 3 months, targets 99.5% uptime and spills over to Standard when you exceed it.

The catch: not every model offers every tier. On 22 October 2026 the model cards for Claude Haiku, Sonnet and Opus 5.5 list Standard only. Flex and Priority exist for Nova 2 Lite and several open models; Reserved exists for Claude 4.5 and 4.6 models. Check the model card before you plan on a discount.

Reserved is easy to misread as a discount. For Claude Sonnet 4.6 on the Global profile, 1K input tokens per minute reserved for 1 month costs $0.18 an hour. A minute of 1,000 tokens for an hour is 60,000 tokens, so $0.18 is exactly $3 per million: the on-demand price at 100% use. The 3-month term takes 10% off. You are paying for guaranteed capacity, so reserve only the traffic you are sure to send every minute and let the rest overflow to Standard.

Prompt caching does most of the saving

When the start of a prompt repeats (a long system prompt, tool definitions, a document users ask about), Bedrock can serve it from the prompt cachePrompt CachingLetting the API keep the processed start of a prompt for a few minutes, so later requests that begin the same way are cheaper and quicker.. A cache read on Sonnet 5.5 costs $0.10 per million instead of $2, a twentieth. Since 7 October 2026 the Sonnet 5.5 read price is half what it was.

The rules: mark up to 4 cache checkpoints in the prompt; the prefix before a checkpoint must be at least 512 tokens on the Claude 5.5 models; the cache lives 5 minutes (or 1 hour, at a higher write price) and each hit resets the clock. Anything that changes the prefix, even a timestamp at the top of the system prompt, is a miss. Our prompt caching checklist covers the usual reasons a cache never hits.

A worked month: a support assistant on Sonnet 5.5

20,000 requests a day, each with a 5,000-token system prompt and help-centre extract that repeats, 1,000 tokens of conversation that does not, and a 400-token answer. Over 30 days that is 600,000 requests:

• Cached input: 5,000 × 600,000 = 3 billion tokens × $0.10 = $300
• Fresh input: 1,000 × 600,000 = 600 million × $2 = $1,200
• Output: 400 × 600,000 = 240 million × $10 = $2,400

$3,900 a month on the Global profile, $4,290 on the US profile. Without caching, the same traffic costs $7,200 for input alone, $9,600 in all. Move it to Haiku 5.5 and the month is about $210, which is why the model choice is usually the biggest lever of all (our Opus vs Sonnet vs Haiku guide helps decide).

Notice that output is $2,400 of the $3,900. Once caching is on, shorter answers save more than anything you do to the prompt.

Try itWhat a month on Amazon Bedrock costs

Model

·Claude Haiku 5.5·Claude Sonnet 5.5·Claude Opus 5.5·Amazon Nova 2 Lite

GlobalGeo (US, EU…)

Requests a day

20,000

Input tokens a request

6,000

Input read from the cache

80%

Output tokens a request

400

Fresh input at $2/M$1,440

Cached input at $0.10/M$288

Output at $10/M$2,400

A month: $4,128Standard tier only

Batch and Provisioned Throughput

Batch inference reads a file of prompts from S3 and writes the answers back, at half the on-demand price, for models that support it. The newest Claude models do not: on 22 October 2026 batch shows N/A for Haiku, Sonnet and Opus 5.5, while Claude Opus 5 and Sonnet 4.6 still list batch prices ($2.50 and $12.50, $1.50 and $7.50). If a half-price overnight run on a 5.5 model matters, Anthropic's own Message Batches API offers it.

Provisioned Throughput buys model units by the hour. Nova 2 Lite costs $60.50 an hour per unit with no commitment, $55 on a 1-month term and $30.25 on a 6-month term, roughly $22,000 a month even at the cheapest rate. You need it to serve a fine-tuned model, and for Claude you arrange it through your AWS account team. Billing runs until you delete it, used or not.

What to do

1. Use the Global profile unless a rule says otherwise.
2. Put everything that repeats at the top of the prompt and add cache checkpoints.
3. Cap answer length; output is usually the largest line.
4. Check the model card for Flex and batch before planning on them.
5. Reserve capacity only for availability, never to save money.

Watch your quota too: Bedrock counts output tokens of newer Claude models several times against tokens per minute, as LLM API rate limits explains. To compare the same job across makers, the cost calculator prices it on every current model. In Cost Explorer, Claude charges appear under the model provider through AWS Marketplace, not under Bedrock, so include both in your reports.