AWS Bedrock Pricing 2026: Per-Token Rates, Service Tiers, Batch and Reserved Explained

Jerzy Kopaczewski 10 November 2026 Updated 21 November 2026 15 min read
Contents

AWS Bedrock Pricing 2026: Per-Token Rates, Service Tiers, Batch and Reserved Explained

AWS Bedrock pricing is not one price list. You pay per million input and output tokens at a rate set for each model, then adjust it with a service tier (Standard, Priority, Flex or Reserved), Batch, prompt caching and the type of cross-Region inference you use. Features such as Knowledge Bases and Guardrails are billed on top. The part most guides skip: AWS publishes tier prices for some model families and not for others, so you need to read the row for your model and measure before you plan. This guide uses rates read from the AWS pricing page on 21 November 2026.

By Jerzy Kopaczewski, Devopsity (AWS consulting partner). All rates are per 1 million tokens, US East (Ohio), Global cross-Region inference, Standard tier, unless a table says otherwise. Check the AWS Bedrock pricing page before you commit: rates change.

New to the service? Start with What is Amazon Bedrock. This page is the pricing deep dive.

 

How AWS Bedrock pricing works

A Bedrock bill is the sum of several meters. The model rate is the biggest, and it is rarely the only one.

  • Input and output tokens, by model. Output tokens cost more than input tokens on every model below, often five times more.
  • Service tier. Standard is the default. Priority costs more for faster responses, Flex costs less for slower ones, and Reserved is a capacity commitment. You pick the tier per request.
  • Batch. For select models, AWS prices batch inference 50% below on-demand.
  • Prompt caching. Cached input tokens bill at a cache-read rate, and tokens written to the cache can bill above the normal input rate.
  • Cross-Region inference. Global profiles are the base rate. Geographic profiles add about 10% on the models checked here.
  • Separately priced features. Knowledge Bases, Guardrails, Model Evaluation, Data Automation, Intelligent Prompt Routing, Prompt Optimization and Web Search each have their own pricing.

 

Standard rates side by side

Standard tier, US East (Ohio), Global cross-Region inference, except DeepSeek (US East in-region). The last column prices one example workload: 500 million input tokens and 100 million output tokens a month, no caching.

ModelInputOutputExample month
Claude Haiku 5.5$0.10$0.50$100
Amazon Nova 2 Lite$0.30$2.50$400
DeepSeek v3.2$0.62$1.85$495
Amazon Nova 2 Pro (Preview)$1.25$10.00$1,625
Claude Sonnet 5.5$2.00$10.00$2,000
Claude Opus 5.5$4.00$20.00$4,000

Three things stand out. Claude Haiku 5.5 is 20 times cheaper than Claude Sonnet 5.5 on both input and output, and Sonnet 5.5 is half the price of Opus 5.5. Haiku 5.5 is also ten times cheaper than Haiku 4.5 ($1.00 input, $5.00 output), so moving to a newer generation can lower the bill. And list price is not value: these numbers say nothing about which model does your task well enough. Pick the cheapest model that passes your own evaluation, then optimise around it.

For the cost of running models yourself instead, see our Bedrock vs SageMaker vs self-hosted EKS comparison.

 

Bedrock service tiers: Standard, Priority, Flex and Reserved

AWS offers four service tiers. You choose per request with the service_tier parameter (reserved, priority, default or flex). If you leave it out, the request runs on Standard.

TierWhat you getHow it is priced
StandardThe default for everyday production trafficBase rate
PriorityRequests served ahead of Standard and Flex, with faster responsesPremium on the base rate, no reservation
FlexSlower, best-effort processing for work that can wait; handled after Standard under heavy loadDiscount on the base rate
ReservedReserved capacity with a 99.5% uptime target; overflow goes to StandardFixed hourly price per 1K tokens per minute, 1 or 3 month term

Priority, Standard and Flex share one on-demand quota per model. Reserved capacity is a separate pool. That has a practical consequence: a large Flex job can consume the quota your Standard traffic needs.

What the discounts and premiums actually are

For the Amazon Nova models, AWS publishes every tier. In US East (Ohio), Global cross-Region inference:

Model (input / output)StandardPriorityFlexBatch
Nova 2 Lite$0.30 / $2.50$0.525 / $4.375$0.15 / $1.25$0.15 / $1.25
Nova 2 Pro (Preview)$1.25 / $10.00$2.1875 / $17.50$0.625 / $5.00$0.625 / $5.00

On these models Priority is 1.75 times Standard, and Flex and Batch are half of Standard. AWS’s pricing page states the same pattern in footnotes for several other providers, for example DeepSeek and the OpenAI open-weight models. Two useful consequences:

  • Flex gives you the Batch price on a synchronous call wherever both exist. You keep a normal request and response and still pay half.
  • Do not assume a flat multiplier. Nova 2 Omni (Preview) text is rounded differently (Priority $0.50 instead of $0.525), and the older Nova Micro and Nova Lite list Batch but no Priority or Flex rows at all. Read the row for your model.

Claude is different

The Anthropic section of the pricing page lists Standard rates, Batch rates for older models and cache rates. It does not publish separate Priority or Flex tables, and AWS’s service-tiers page describes Flex only as a discounted standard rate with supported models to be checked in the documentation. So we do not tell you that Claude Flex is half price or that Claude Priority is 1.75 times Standard. Confirm tier availability and rates for your exact Claude model with AWS before you plan around them. A blanket “move everything to Flex” plan can fail on the models that matter most.

 

A tier mix that saves money (and one that does not)

Take the Nova 2 Lite workload from above: 500 million input and 100 million output tokens a month. All Standard costs $400. All Flex costs $200. All Priority costs $700.

MixMonthly costVersus all Standard
100% Standard$400baseline
60% Standard, 40% Flex$320-20%
60% Standard, 30% Flex, 10% Priority$370-7.5%

A 10% Priority slice adds $30. A 30% Flex slice saves $60. Priority is worth buying only for the traffic where a slow response costs you more than the premium. The default plan that holds up: customer-facing requests on Standard, background and evaluation work on Flex, and Priority only for the narrow slice that needs it. Measure the split before you commit to it.

 

Batch versus Flex

AWS prices batch inference 50% below on-demand for select models. Batch is asynchronous: you submit a file, results land in S3, and jobs are best effort. On-demand requests take priority over batch jobs, a job times out after 24 hours by default (configurable up to 168 hours), and completion is not guaranteed. Where a model offers Flex, Flex gets you the same discount without the file-based workflow, so Batch earns its place for bulk jobs and for models that have no Flex row.

One catch from the Claude table: Batch rates are listed for older models such as Sonnet 4.6, Opus 4.6 and Haiku 4.5, and shown as not available for the 5.5 line. For the newest Claude models, model choice and caching are the levers, not Batch.

 

Reserved tier: a capacity purchase, not a discount

AWS publishes Reserved rates for some Claude models, priced per hour per 1,000 tokens per minute (TPM). US East (Ohio), Global cross-Region inference:

ModelInput, 1 monthOutput, 1 monthInput, 3 monthsOutput, 3 months
Claude Haiku 4.5$0.06$0.30$0.054$0.27
Claude Sonnet 4.6$0.18$0.90$0.162$0.81
Claude Opus 4.6$0.30$1.50$0.27$1.35

Here is the arithmetic that most guides leave out. One thousand tokens per minute, used flat out for an hour, is 60,000 tokens. At Claude Sonnet 4.6’s Standard input rate of $3.00 per million, 60,000 tokens cost $0.18. That is exactly the 1-month Reserved hourly rate. The same holds for Sonnet’s output ($15 per million gives $0.90), and for Haiku 4.5 and Opus 4.6. In other words, on the published rates, 1-month Reserved costs the same as Standard on-demand only if you use 100% of what you reserved. The 3-month rate is 10% lower, so it breaks even at 90% utilisation. Below that, every unused token is paid for anyway.

So Reserved is not a discount on these rates. It buys guaranteed capacity, a 99.5% uptime target and protection from throttling, and overflow still bills at Standard. That is a good trade for steady, customer-critical traffic. It is a poor one for spiky traffic. AWS’s own capacity guidance quotes savings of 30 to 50% for steady high-volume load. We could not reproduce that from the published rates for these models, so ask your AWS account team what applies to your terms.

There is also a minimum. Reserved capacity starts at 100,000 input TPM and 10,000 output TPM, and access goes through your AWS account team. At 730 hours a month on 1-month rates, that minimum is about $6,570 a month for Haiku 4.5, $19,710 for Sonnet 4.6 and $32,850 for Opus 4.6 (about 10% less on 3-month terms). Below that scale, Reserved is not on the table. Two more traps: billing continues until the reservation is deleted with your account manager, and when you size it you count cache-write tokens as well as input tokens. We could find no Reserved rates for the 5.5 models, so check availability before you plan.

 

Prompt caching

Caching reduces the price of input you send repeatedly, such as a long system prompt or a shared document. There are three input rates: standard input, cache write and cache read. Cache entries live on the server that handled the request, so a request routed elsewhere can write a new entry instead of reading one.

ModelInput5-min cache write1-hour cache writeCache read
Claude Haiku 5.5$0.10$0.125$0.20$0.01
Claude Sonnet 5.5$2.00$2.50$4.00$0.10
Claude Opus 5.5$4.00$5.00$8.00$0.20

For Sonnet 5.5 a 5-minute cache write costs 1.25 times the input rate and a read costs 0.05 times, so a cached prefix pays for itself on its second use. AWS notes that Sonnet 5.5 cache-read pricing fell by 50% on 7 October 2026. For Amazon Nova, AWS states that cache reads cost 75% less than the on-demand input price. Caching support, minimum prompt length and cache lifetime vary by model, and implicit caching is best effort, so confirm in the model card.

 

Cross-Region inference: Global or Geo

Cross-Region inference routes requests across Regions for throughput. AWS adds no routing charge: you pay the price of the Region you call. Global profiles carry the base rate. Geographic and in-region pricing is higher on the models checked here, exactly 10% for Claude Sonnet 5.5 ($2.20 input and $11.00 output instead of $2.00 and $10.00) and for Nova 2 Lite ($0.33 and $2.75 instead of $0.30 and $2.50). The trade-off is data residency: a Global profile can process requests in any supported commercial Region. Some older Claude models show the same price in both columns, so read the row.

 

Measure before you plan

Tier discounts are not uniform, so test each plan on your own traffic before you rely on it.

  1. Send a test request with service_tier set to the tier you want.
  2. Read the resolved tier. The tier that actually served the request appears in the API response, in CloudTrail, and in the ResolvedServiceTier CloudWatch dimension. Alarm when it differs from what you asked for.
  3. Check the billed line in Cost Explorer after a day of test traffic, and compare it with your estimate.
  4. Split cost by workload with application inference profiles. Each profile tracks usage and cost for one model or cross-Region profile, and supports tags for cost allocation.
  5. Watch throttling. If throttling hits users, upgrade the tier for that path. If not, do not pay for it. Our runbooks cover 429 throttling and the tokens-per-day limit.

 

Cost traps Devopsity sees

  • A forgotten Reserved reservation. It bills monthly until it is deleted through your account manager.
  • Flex jobs eating Standard quota. The on-demand quota is shared across Priority, Standard and Flex.
  • Under-sized Reserved capacity. Cache-write tokens count toward your tokens per minute.
  • Agent loops. An agent that plans, calls tools and rereads context multiplies tokens. Cost the loop, not the prompt. See AgentCore pricing.
  • Assuming one multiplier. Read each model’s row. The 1.75 and 0.5 pattern holds for Nova and several providers, and is not published for Claude.
  • Paying for retries. Failed or repeated calls that still consume tokens add up, so set a retry budget.

 

Want a second opinion on your Bedrock cost model before you commit to a tier?

Book a free 30-min call

 

Frequently asked questions about AWS Bedrock pricing

How does AWS Bedrock pricing work?

You pay per million input and output tokens at a rate set for each model. A service tier (Standard, Priority, Flex or Reserved), Batch, prompt caching and the type of cross-Region inference then adjust the price, and features such as Knowledge Bases and Guardrails are billed separately. AWS publishes the rates on its Bedrock pricing page.

Is Amazon Bedrock free?

No. Bedrock is billed by usage. New AWS accounts may receive promotional credits that can offset it, so check the current AWS Free Tier terms.

How much does Bedrock cost per million tokens?

It depends on the model. On 21 November 2026 AWS listed Standard rates from $0.10 input and $0.50 output per million tokens for Claude Haiku 5.5, to $4.00 and $20.00 for Claude Opus 5.5, in US East (Ohio) with Global cross-Region inference. Output tokens cost more than input tokens.

What are the Amazon Bedrock service tiers?

Four tiers: Standard (the default), Priority (faster, at a premium), Flex (slower, at a discount) and Reserved (reserved capacity on a 1 or 3 month term). You select one per request with the service_tier parameter, and support varies by model.

Is Bedrock Flex cheaper than Standard?

Where a model offers it, yes. AWS lists Flex at half the Standard price for Amazon Nova and several other providers, but does not publish Flex rates for Claude. Check your model's row, and test before you plan around it.

Is Bedrock Reserved tier worth it?

Only for steady, high-volume, customer-critical traffic. On the published Claude rates, 1-month Reserved costs the same as Standard on-demand at 100% utilisation, and 3-month breaks even at 90%, so it buys capacity and uptime rather than a discount. Minimum commitments run into thousands of dollars a month.

 

Summary

Bedrock pricing is a per-model rate adjusted by tier, Batch, caching and Region type. Three things to remember:

  1. Read your model’s row. Tier prices are published for some families and not others.
  2. Reserved buys capacity, not a discount, on the published Claude rates.
  3. Measure first. Test the tier, confirm the resolved tier, and check the bill before you commit.

Devopsity helps teams model Bedrock cost before rollout and keep it under control afterwards, through our FinOps and AWS consulting work.

Jerzy Kopaczewski

Planning a Bedrock workload?

Book a free 30-minute call with Devopsity. We help teams pick models and tiers, model token and retrieval costs before rollout, and avoid the Reserved and quota traps above.

Book a call
AWS Amazon Bedrock FinOps GenAI LLM pricing Claude Amazon Nova