AWS Bedrock cost calculator

How Bedrock billing actually works

Amazon Bedrock is a gateway, not a single model. You call Claude, Amazon's own Nova family, Meta's Llama, Mistral or Titan through one API, and each is metered the same way: a rate per million input tokens (everything you send — system prompt, context, the user's message) and a separate, usually higher, rate per million output tokens (what the model generates). There's no monthly minimum on on-demand; your bill is purely tokens × rate. The subtlety most estimates miss is that output is where the money goes — it's priced two to five times higher than input on the Claude and Nova tiers, so a chatty model that writes long answers costs far more than its input rate suggests.

This calculator makes that visible. Enter the tokens a typical request sends and receives, and your monthly request count, and it splits the bill into an input share and an output share so you can see which one to attack. Then it prices the same workload across the whole Bedrock catalog and ranks them, because the cheapest model for a summariser that reads a lot and writes a little is not the cheapest model for a generator that writes long output. The headline "per 1M" number rarely tells you the winner — your token mix does.

The three levers that cut a Bedrock bill

1. Model choice. Dropping from a frontier Claude to Nova Lite or Llama for the requests that don't need deep reasoning can cut cost by 90%+ — the table shows exactly how much for your mix. Route hard prompts to the big model and easy ones to a cheap one; size the split with the model routing calculator. 2. Batch inference. If the work isn't real-time — overnight enrichment, bulk classification, evals — Bedrock's batch mode runs it at roughly half price. Flip the pricing mode above to see it. 3. Prompt caching. When many requests share a large fixed prefix (a long system prompt, a document, few-shot examples), caching that input discounts it by up to 90%; set the cached-input percentage to model the saving, and dig deeper with the prompt caching calculator.

Bedrock vs the direct API

For the same Claude model, per-token list prices on Bedrock and the direct Anthropic API are typically the same or within a rounding error. So the choice is operational, not financial: pick Bedrock to keep inference inside your AWS account, IAM and VPC, to consolidate billing, and to reach Nova, Llama, Mistral and Titan through one integration. Pick the direct API for the newest models on day one and the simplest setup. If cost is the whole question, compare the winner here against the cheapest LLM API across providers before you commit.

How to use it

1. Pick a model and enter your monthly request volume.
2. Enter typical input and output tokens per request (a rough average is fine).
3. Choose on-demand or batch, and set the share of input that's cacheable.
4. Read the monthly cost and per-request cost, then scan the table to see if a cheaper model covers the same job.

Common mistakes

Estimating on input tokens only. Output is the expensive half — always include a realistic output length. Using averages that hide long replies. If some responses are 5× longer, your real bill runs above the average; model the worst case too. Forgetting the free levers. Batch and caching are set-once savings that many teams never turn on. Ignoring surrounding costs. Data transfer, storage for RAG, and Provisioned Throughput commitments sit outside per-token pricing — budget for them separately.

FAQ

Estimate only. Bedrock prices differ by region and change — verify current rates with AWS before budgeting.