Skip to content

How Amazon Bedrock pricing works, without the engineering

Amazon Bedrock bills on-demand use by tokens and Provisioned Throughput by model unit per hour. Learn what drives the bill and how to estimate it.

6 min readBy AI Guru AcademyFacts last checked

By AI Guru Academy

In short

Amazon Bedrock bills on-demand use by the tokens a model processes, with input and output tokens listed separately. Provisioned Throughput is a different structure: you buy capacity in model units and are billed per model unit per hour. Cost therefore follows prompt size, answer length and request volume for on-demand use, and follows the capacity you reserve for Provisioned Throughput. This article describes the structures only and quotes no prices, because rates change and vary by model.

How does on-demand Amazon Bedrock pricing work?

On-demand use is charged by tokens processed, and the pricing page lists input tokens and output tokens as separate line items. A token is a small piece of text that the model reads or writes. Longer text means more tokens.

The pricing page states its on-demand rates per 1 million tokens, and the tables shown there are for the Standard tier. Other options, such as the Priority and Flex tiers and batch inference for some models, are priced differently, so check which one you will use. You pay for what you use, with no capacity to buy first. If nobody calls the model for a week, on-demand usage for that week is zero.

Two consequences follow for a manager. First, the bill is a product of three things: how much you send, how much the model writes back, and how many times you do it. Second, the rates differ between models, so the choice of model is itself a cost decision.

How does Provisioned Throughput pricing differ?

Provisioned Throughput is billed hourly for the capacity you purchase, and billing continues until you delete it, so you pay whether or not you use all of it. AWS describes it as a way to provision a higher level of throughput for a model at a fixed cost. A model unit is the unit in which that capacity is sized: each one delivers a set number of input and output tokens per minute for the chosen model.

The hourly price depends on the model, the number of model units and the commitment term you choose. A longer commitment lowers the hourly price, and with a commitment you cannot delete the Provisioned Throughput until the term ends. Rates are not published for every model; for some, the pricing page directs you to the AWS account team. That means you cannot always work out a Provisioned Throughput cost from the public page alone.

How does Provisioned Throughput pricing differ?
StructureWhat you pay forCost moves with
On-demandTokens processed, input and output priced separatelyPrompt size, answer length, request volume
Provisioned ThroughputModel units, billed per hourNumber of model units, hours held, commitment term

What actually drives an on-demand bill?

Three levers drive it: input size, output size and volume.

  • Prompt size. Every instruction, example and retrieved document you place in a prompt is billed as input tokens.
  • Output length. Answers are billed as output tokens, priced separately from input. Models typically support an inference parameter that limits response length, and the available parameters and their defaults depend on the model.
  • Volume. Ten thousand requests cost roughly ten thousand times one request of the same shape.

A very long prompt costs more because every token in it is billed as input. Each model also has a context window that limits how much it can handle in one request; check the model's documentation for its limit.

Illustration (invented, not a real company). A fictional training firm builds an assistant that answers staff questions from policy documents. Version one pastes a whole 40-page policy into every question. Version two sends only the three relevant paragraphs. Both give the same answer quality in the firm's tests, but version two sends far fewer input tokens per question. With the same number of questions, version two costs less on an on-demand basis. The saving comes from prompt design, with no change of vendor or model.

When does Provisioned Throughput make sense?

Consider Provisioned Throughput when you need reserved capacity at a fixed cost, and stay on-demand when your usage is uncertain or uneven. This is a decision rule, and the documents above support its two halves: on-demand token charges stop when nobody calls the model, while Provisioned Throughput charges hourly for what you purchased until you delete it.

A practical checklist before choosing a fixed commitment:

  1. Do you have measured, steady usage, not a forecast made before launch?
  2. Do you need capacity that is set aside for you at a fixed cost?
  3. Can you accept paying for the capacity during quiet hours?
  4. Have you confirmed the rate for your chosen model with AWS, since it is not published for every model?

If you cannot answer yes to the first three, on-demand is the safer starting point, because there are no token charges when the model is not being called while the workload is still being learned.

How is this different from instance-based and seat-based services?

Bedrock on-demand charges by usage volume. Other AWS services charge by a different unit, and comparing units is a common exam and boardroom theme.

  • Amazon SageMaker AI hosting. Real-time hosted endpoints are charged for usage of the instance type you choose. Serverless endpoints differ: they are billed for the compute capacity used to process requests, plus data processed, not by instance time.
  • AWS Quick (formerly Amazon Quick). Priced mainly per user, so the bill largely follows headcount. Some plans also carry an account-level infrastructure fee and usage-based overage charges.
  • Amazon Bedrock on-demand. Priced by tokens, so the bill follows how much the model is used.
How is this different from instance-based and seat-based services?
Service typeBilling unitBill follows
Bedrock on-demandTokensUsage volume
Bedrock Provisioned ThroughputModel unit per hourReserved capacity
SageMaker AI real-time endpointInstance usageInstance type and running time
AWS QuickMainly seat (per user)Mainly number of users

The practical point is that a pilot with ten users says little about a rollout to a thousand if the service bills by usage, and says much more if it bills mainly by seat.

How do I estimate and track the cost?

Use the AWS Pricing Calculator before launch and AWS Cost Explorer after. The Pricing Calculator is a free web-based planning tool for creating cost estimates. You can use it to model a solution before building it or to plan a change to existing usage. Its estimates reflect only the information you enter and exclude taxes, so they are estimates and not quotes.

Cost Explorer lets you view and analyse your costs and usage over time. It can group and filter costs, for example by service or usage type, which helps you see which part of a bill is growing. Historical views cover a limited period, so check the AWS page for what it shows.

A simple routine: estimate with realistic prompt and answer lengths, launch small, then compare Cost Explorer against the estimate and adjust the prompt design or model choice.

Key takeaways

  • On-demand Bedrock use is billed by tokens processed, with input and output tokens priced separately.
  • Provisioned Throughput is billed per model unit per hour for capacity you buy, used or not.
  • On-demand cost follows prompt size, answer length and request volume.
  • Model choice changes cost, so compare models on quality and price together.
  • Seat-based and instance-based services bill on different units from token-based ones, so compare the unit before comparing totals.
  • Estimate with the AWS Pricing Calculator, then track actual spend with Cost Explorer.

FAQ

Is Amazon Bedrock priced per user?

No. On-demand use is priced by tokens processed, so the bill follows usage. Some other AWS services, such as AWS Quick (formerly Amazon Quick), are priced mainly per user.

What is a model unit?

It is the unit in which Provisioned Throughput capacity is sized and billed hourly. Each model unit delivers a set throughput for a specific model. The rate depends on the model and the commitment term, and AWS does not publish a rate for every model.

Do input and output tokens cost the same?

They are listed as separate prices on the pricing page, so you should estimate them separately. Check the page for the model you plan to use.

Does the Pricing Calculator give me a guaranteed price?

No. It produces estimates from the inputs you enter, and it excludes taxes. Your actual bill depends on real usage.

Sources

Last checked 2026-11-16.

AI Guru Academy is an independent training provider and is not affiliated with or endorsed by Amazon Web Services. AWS service names are trademarks of Amazon.com, Inc. or its affiliates.

Next step: take the free AIB-C01 diagnostic at /diagnostic, then see the full course at /courses/aib-c01-exam-prep.

How Amazon Bedrock pricing works, without the engineering | AI Guru Academy