Amazon Bedrock in Practice

How Bedrock Claude Charges Appear on Your AWS Invoice

The first Bedrock invoice after a Claude launch raises the same question in every finance team: what are all these token line items, and why are some of them five times more expensive than others?

Claude 3P 101 · Updated July 2026 · Unofficial guide

Claude on Amazon Bedrock is billed through your normal AWS invoice — there is no separate Anthropic invoice, and marketplace list prices for Claude match Anthropic's first-party list prices. Within the Amazon Bedrock service section of the bill, charges break down into line items keyed by usage types: metering strings that identify the region, the model, and the unit being charged. This article explains the categories those line items fall into, so the invoice reads as a story rather than a wall of codes. The exact usage-type strings are best read from your own bill or Cost and Usage Report — they are account-visible ground truth, and this guide deliberately describes categories rather than reproducing code strings.

The core pair: input tokens and output tokens

On-demand inference — the default pay-as-you-go mode — is charged per input token and per output token consumed, at per-model rates. Every Claude model therefore produces at least two recurring line items: one metering the tokens you sent (prompts, documents, tool definitions, conversation history) and one metering the tokens Claude generated (including any extended-thinking tokens, which bill as output). The rates differ substantially: at list, Claude Opus 4.8 is $5 per million input tokens and $25 per million output tokens; Claude Sonnet 5 is $3/$15 at standard pricing (with promotional $2/$10 pricing through August 31, 2026, per AWS's pricing page); Claude Haiku 4.5 is $1/$5. That 5x input/output spread is why two applications with identical request volume can have wildly different bills — verbose output is the expensive direction.

Cache writes and cache reads

If your application uses prompt caching, expect additional token line items, because Bedrock meters cache activity at its own rates: a cache-write rate (a premium over base input) and a cache-read rate (a tenth of base input). A healthy caching setup shows modest cache-write volume and large cache-read volume. The inverse pattern — heavy writes, few reads — on an invoice means you are paying the caching premium without harvesting the discount, which is a code problem (an unstable prompt prefix) surfacing as a billing symptom.

Other line items you may see

Line item categoryWhat it means
Batch inference tokensAWS's S3-based batch mode, priced 50% below on-demand for supported models
Provisioned Throughput hoursFixed hourly billing per Model Unit, with no-commitment, 1-month, or 6-month terms — appears as time-based charges, not token charges
Regional-endpoint premiumRegional (single-region) endpoints price 10% above global endpoints for Claude Sonnet 4.5 and later models; the premium is embedded in the regional rates
Adjacent AWS servicesCloudWatch Logs, S3, and similar charges from invocation logging or batch I/O bill under those services, not Bedrock

Region also shapes the invoice's structure: usage metered in different regions appears under different regional line items, and with cross-region inference, pricing is based on the region from which you call the inference profile, with no additional routing cost.

Reconciling the bill against reality

To sanity-check a line item, multiply your measured token volume by the published rate. Token volume comes from your application's response metadata (every response reports input and output token usage) or, in aggregate, from model invocation logging, which records per-request modelId and input/output token counts for calls on the bedrock-runtime endpoint. Rates come from AWS's Bedrock pricing page. If the multiplication doesn't match the invoice, the usual suspects are: a model-generation change (Claude Opus 4.7-and-later models use a tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6-and-earlier — the same traffic literally becomes more tokens), promotional pricing expiring, an unnoticed shift between global and regional endpoints, or caching behavior changing.

Rule of thumb: read the invoice as four questions — which model, which direction (input/output/cache), which region and endpoint type, and on-demand or provisioned. Every Bedrock Claude line item is an answer to those four.

Where to go next

Analyzing Claude costs in Cost Explorer covers the interactive view of the same data; the pricing explainer covers token billing from first principles; and invoice reconciliation goes deeper on the finance workflow.

Sources