AWS Bedrock Pricing and How to Estimate Your Bill

0
AWS Bedrock Pricing

Fifty thousand requests on Amazon Bedrock could cost $12 in model fees. The same number of requests, with the same amount of text, could cost $675 on another model. Both figures can be right. And if each request searches an internal document library before it answers, you’ll have another line on the bill.

That’s what makes AWS Bedrock pricing tricky to estimate from a rate card alone. A customer might type, “Can I return this?” Your app might send the entire conversation, two pages of return policy, and instructions for how to handle exceptions. AWS meters what your app sends and what the model generates. To get a believable estimate, start with one real request and follow it through your app.

Table of Contents

  1. What Bedrock Charges For
  2. What 50,000 Requests Cost
  3. When It Pays to Wait for a Response
  4. Where the Extra Charges Come From
  5. How to Cut the Cost per Request
  6. Build a Budget From Real Traffic

What Bedrock Charges For

For on-demand text generation, Bedrock generally charges for input and output tokens at separate rates. Tokens are the pieces of text a model processes. A token might be a word, part of a word, or punctuation. Your input can include system instructions, the user’s message, earlier turns, and any reference material the app attaches. The output is what the model generates. If the app sends a generated answer back as part of the next turn’s history, that text counts toward the next request’s input.

Check the rate unit before you do the math. AWS lists prices per 1,000 tokens in some tables and per million in others. For rates quoted per million, divide your input token total by one million and multiply by the input rate. Do the same for output, then add the two amounts. Image and audio workloads can use different billing units, so use the rate card for the specific model and media type you’re sending.

The routing option also matters. In-Region inference processes requests in one AWS Region. Geographic cross-Region inference can route them within a defined area such as the US or EU. Global inference can route across supported commercial Regions worldwide. Check both the price and where your data can be processed before choosing a profile.

What 50,000 Requests Cost

Say you run a support assistant that handles 50,000 requests a month, with one model call per request. Each call averages 2,000 input tokens and 500 output tokens. That’s 100 million input tokens and 25 million output tokens in a month. Here’s how the model fees work out at selected USD rates checked November 4, 2026.

Amazon Nova Lite in US East (Ohio) costs $0.06 per million input tokens and $0.24 per million output tokens. The example comes to $12.00 a month.

Mistral Large 3 in US East (Ohio) costs $0.50 per million input tokens and $1.50 per million output tokens. That comes to $87.50 a month.

Claude Sonnet 4.6 with Global Cross-Region routing costs $3.00 per million input tokens and $15.00 per million output tokens. That comes to $675.00 a month.

Look at the Claude example. The input costs $300 and the output costs $375, even though the model generates only a quarter as many tokens as it receives. If a shorter reply still does the job and you get the average down from 500 to 200 output tokens, you save $225 a month at the same traffic level. That’s worth testing on simple questions where people just want an answer and a next step.

These examples hold token counts steady so you can see the effect of the rates. For your own comparison, give each model the same representative tickets and count how many it resolves correctly. A low per-call price can lose its appeal if the app regularly calls a second model to repair the first answer or sends the work back to a person.

When It Pays to Wait for a Response

If a customer is waiting in a chat window, on-demand Standard is a reasonable baseline. For supported models, Priority costs more and gives requests priority over Standard and Flex. Flex offers lower rates for work that can tolerate slower responses. A background task that drafts notes for tomorrow’s queue is a better candidate for Flex than the reply a customer is waiting to read.

Batch inference is a separate option for work you can submit as a job. Imagine a team summarizing last month’s call transcripts. It can upload the inputs to Amazon S3, start the job, and pick up the results later. AWS lists batch rates 50% below on-demand rates for select models. Check availability for your model and Region, along with the job limits and timing. A backfill with a flexible deadline is a natural fit.

If predictable capacity is the problem, AWS’s Reserved tier is another pricing conversation. It uses a commitment for input and output capacity, and you arrange it through your AWS account team. Compare it with what you’d actually use over a full week, including quiet nights and weekends. A busy hour gives you a useful peak figure and a poor picture of average demand.

Where the Extra Charges Come From

Go back to that returns chatbot. If it uses a Bedrock Managed Knowledge Base to find policy excerpts, storage and retrieval have their own charges. At AWS’s published rates, indexing 5 GB of raw data costs $25 a month. Making 50,000 standard retrieval calls adds $50. That’s $75 for the knowledge base before the model writes a single reply. With a self-managed setup, price your own vector store, embedding model, and any optional reranker separately.

Guardrails can add a charge for each policy you use. Say you apply the standard text content filter to both the input and the output of every request, and each piece of text fits within one text unit of up to 1,000 characters. At 50,000 requests, that’s 100,000 text units. At $0.15 per 1,000 units, the filter adds $15. Longer messages can use several units, and enabling another paid policy adds its own charge.

There’s also a difference between a user action and a model call. One click might trigger a call to decide what to look up and a second call to write the answer. If the app repeats either step when the response comes back in the wrong format, include that extra call in the estimate. A simple trace of what happens after the click can explain a surprising bill.

How to Cut the Cost per Request

I’d look at the prompt before changing models. If the app sends six policy excerpts for every question and people usually need one or two, see what happens when you retrieve fewer. Keep the passages that cover tricky exceptions. Test questions like “Can I return an opened item bought on sale?” before you trim further. The cheaper prompt still has to help someone make the right decision.

Prompt caching can help when supported models see the same long instructions or reference material over and over. A cache hit can cost less than processing that input again at the standard rate. Writes can cost more, and entries expire, so a long shared prefix that gets reused every minute is a better fit than a policy document someone opens once a day. Measure cache reads and writes on real traffic before counting the savings.

Build a Budget From Real Traffic

Run a small pilot with a mix of quick questions, long conversations, and the awkward cases that take several steps. Where supported, the free CountTokens API can count inputs before you send them for inference. Once requests run, use the model’s usage data or Bedrock’s CloudWatch metrics to measure actual input and output. Track cache reads and writes separately if you use caching. Then add retrieval, filtering, and any other services the pilot used.

Put those numbers into a quiet month and a busy month. A holiday sale might double traffic and bring in longer refund conversations at the same time. You can set AWS Budgets alerts for actual or forecast spending, and use application inference profiles with cost allocation tags to help separate model spend by workload.

The number I’d keep on the dashboard is cost per resolved request. If it rises, you can look for the change that drove it, whether that’s longer replies, more lookups, or more tickets that need a second pass. Once you’ve measured a few real requests from start to finish, the pricing page becomes a tool you can use to plan the next month’s bill.

Leave a Reply

Your email address will not be published. Required fields are marked *