How the calculator prices a workload
It takes AWS's prices for the model in your Region - per million input, output, cache read and cache write tokens, for every service tier and routing AWS sells it with - and multiplies them by a month of your requests: the requests per day times 730 hours, as AWS bills a month. Every combination the model has a price for becomes a row, and the cheapest comes first. A combination without a price - Claude has no Flex, many models have no Reserved tier - is not offered for that model, so it is not shown.
Third-party models such as Claude are sold through AWS Marketplace, each as its own "(Amazon Bedrock Edition)" service, and the AWS Price List API has their prices next to Amazon's and the open-weight models'. The calculator reads both offers again on every build of this site; the prices shown were published on 2026-09-30.
Amazon Bedrock service tiers
| Tier | Price | For |
|---|---|---|
| Standard | On-demand price | Everyday requests; the default when service_tier is missing or "default" |
| Priority | AWS's pricing page: 75% above Standard | Customer-facing requests that need the fastest responses; served before Standard and Flex |
| Flex | AWS's pricing page: 50% below Standard | Workloads that can handle longer processing times - evaluations, summarization, agent steps |
| Reserved | Per 1,000 tokens per minute per hour, 1- or 3-month term | Mission-critical traffic that cannot tolerate downtime; 99.5% uptime target |
| Batch | AWS: 50% lower than on-demand for select models | Prompts processed asynchronously from JSONL files in S3 |
The calculator uses each model's own prices rather than these percentages: not every model has every tier, and the Price List's prices do not always follow them exactly. Priority, Standard and Flex share the model's on-demand quota - the Bedrock throttling calculator checks whether your traffic fits it.
Prompt caching
Tokens read from the prompt cache are billed at the model's cache read rate, far below input; tokens written to it, for some models at a rate above input (Claude), for others at no extra charge. Tokens not read from the cache are billed as input. The cache's time to live (TTL) resets with each hit, so a prefix used more often than every 5 minutes stays cached; for Claude models with a 1-hour TTL, writing costs more but the prefix survives longer gaps. The calculator counts the prefix of every request as either a hit or a write - your cache hit rate decides whether caching pays off. Prompt caching is not supported with batch inference, and a prefix is cached only from the model's minimum number of tokens per cache checkpoint (1,024 for Claude Sonnet 4.6, 4,096 for Claude Haiku 4.5).
When the Reserved tier pays off
A reservation is priced per 1,000 tokens per minute (TPM) per hour, input and output apart, at least 100,000 input and 10,000 output TPM. A thousand tokens per minute for an hour is 60,000 tokens - and in AWS's price list, a 1-month reservation costs exactly what 60,000 Standard tokens cost, a 3-month one 10% less. So a reservation costs less than Standard only when it is used almost every minute: from 100% of its capacity for 1 month, from about 90% for 3 months. It is a way to buy availability, not a discount. Size it to the busiest minute: tokens above it overflow to Standard, and input tokens written to the cache count against it too.
Frequently asked questions
Is Global cross-Region inference cheaper?
For many models, yes - Claude Sonnet 4.6 costs about 10% less through a global. profile than through a Geo profile, at the price of the Region you call from. A Global profile may route a request to any commercial Region; a Geo profile keeps it in one geography, such as the US or the EU.
Why do some models have no Flex, Priority or batch row?
Not every model supports every tier - AWS's model cards list which ones a model supports, and AWS's price list has prices only for those. The calculator shows what the price list has for the model in your Region.
Will my bill match?
AWS gives the Price List API's prices for informational purposes and charges the price on its Bedrock pricing page where the two differ. Your bill also has what this calculator leaves out - Provisioned Throughput and customization, which the Bedrock fine-tuning cost calculator prices, Guardrails, Knowledge Bases, image and video models. The Bedrock billing decoder reads a real bill per model and meter.
Is what I enter sent anywhere?
No. The calculation runs in your browser. The page only counts that the calculator was used, with the strategy that came out cheapest, never the numbers.
References
Amazon Bedrock pricing
Service tiers for optimizing performance and cost (Amazon Bedrock User Guide)
Prompt caching for faster model inference (Amazon Bedrock User Guide)
Process multiple prompts with batch inference (Amazon Bedrock User Guide)
Global cross-Region inference (Amazon Bedrock User Guide)
AWS Price List (AWS Billing User Guide)