Amazon Bedrock in Practice

Analyzing Claude Costs in AWS Cost Explorer

One advantage of running Claude on Bedrock is that spend shows up in the same tooling as the rest of your AWS bill. Getting from "Bedrock cost $X" to "team Y's summarizer cost $Z" takes a little setup.

Claude 3P 101 · Updated July 2026 · Unofficial guide

AWS Cost Explorer is the console tool for slicing your AWS bill by service, account, region, usage type, and tags. Because Bedrock is a native AWS service billed like any other, Claude usage lands there automatically — no separate vendor portal, no invoice reconciliation against a second system. This article covers how to isolate Claude spend, what the cost dimensions mean, and how to attribute spend to teams and applications. One caveat up front: Bedrock does not support Anthropic's Usage and Cost API, so AWS billing tooling is your cost visibility on this platform.

Step 1: Isolate Bedrock spend

Start by filtering Cost Explorer's Service dimension to Amazon Bedrock. Group the result by region and by linked account to see where usage concentrates. Then group by the usage type dimension — the granular metering strings AWS uses to price each unit of consumption. For Bedrock's on-demand pricing, the units that matter are input tokens and output tokens per model, and with prompt caching in play, cache writes and cache reads are metered at their own rates. The exact usage-type strings encode region, model, and token category; read them in your own account rather than from any third-party list, since they are the ground truth for what you were charged. Grouping by usage type is usually the single most informative view: it separates input from output spend and reveals whether caching is actually being read, not just written.

Step 2: Know what drives each dimension

Interpreting the numbers requires knowing Bedrock's pricing structure for Claude:

Cost driverHow it's billed
On-demand inferencePer input token and per output token consumed, at per-model rates
Prompt cachingSeparate cache-write and cache-read rates per million tokens
Batch inference50% lower than on-demand pricing for supported models
Provisioned ThroughputFixed hourly rate per Model Unit; cheaper hourly with 1- or 6-month commitment
Endpoint choiceRegional endpoints carry a 10% premium over global endpoints (Claude Sonnet 4.5 and later)

Two structural notes explain surprises. Output tokens cost several times more than input tokens (for example, Claude Sonnet 5 lists at $3 input / $15 output per million at standard pricing), so a spike in output-heavy workloads moves the bill disproportionately. And the regional-versus-global endpoint decision is itself a cost lever: AWS cites roughly 10% savings for global cross-region inference relative to geographic routing, with pricing based on the source region of the call.

Step 3: Attribute with tags and accounts

Cost Explorer can only attribute what you've labeled. The standard mechanisms apply to Bedrock as to any AWS service: activate cost allocation tags on the Bedrock-related resources your architecture uses, and enforce a tagging standard (application, team, environment, cost center) in provisioning. Where tagging is too coarse — for example, several applications sharing one account and role — two Bedrock-specific tools close the gap. Model invocation logging records modelId, input and output token counts, the caller's identity ARN, and an optional caller-supplied requestMetadata field per request, which lets you compute per-application token usage from logs and join it to pricing. Alternatively, splitting workloads across member accounts makes the linked-account dimension do the attribution for free. Note that invocation logging currently captures calls through the bedrock-runtime endpoint, not the newer bedrock-mantle endpoint, so mantle-surface workloads need account- or tag-based attribution.

Rule of thumb: account boundaries for coarse attribution, tags for medium, invocation-log token counts for fine. Decide the granularity finance actually needs before building the finest one.

Step 4: Watch it, don't just report it

Once the filters and groupings above give you a clean view, save it as a Cost Explorer report and put an AWS Budget with alerts on Bedrock spend so growth is a notification, not a month-end surprise. Review the usage-type breakdown when anything jumps: the input/output/cache split usually identifies the cause — a longer prompt, a chattier output format, or a cache that stopped hitting — faster than any dashboard total.

Where to go next

See how Bedrock Claude charges appear on the invoice for the line-item view of the same data, cost allocation tags for Bedrock for the tagging setup, and cost alerts and budgets for the alerting side.

Sources