# Amazon Bedrock Guardrail

> Bedrock guardrails bill per text unit evaluated, with each enabled policy type metering separately, so a guardrail is a per-request surcharge on every model call it inspects. ZopNight discovers guardrails via its Bedrock provider on the 6-hour cycle and tracks evaluation cost from Cost Explorer or CUR 2.0 against traffic.

Source: https://zop.dev/integrations/aws/bedrock-guardrail
Updated: 2026-08-19

---

A Bedrock guardrail applies configurable content and safety policies to model calls, billed per text unit evaluated. Guardrail charges scale with every request they inspect, adding a per-call surcharge to inference traffic.

## The inspection surcharge

Guardrails meter on text units, normalized chunks of the input and output they evaluate, and each policy category enabled on the guardrail carries its own per-unit rate: content filters, denied topics, sensitive-information filters, and contextual grounding checks each meter what they inspect. The cost model is therefore a multiplication attached to someone else's traffic: whatever volume the guarded application drives, the guardrail bills a slice of it, per policy, on both the prompt and the response. Enabling one more policy type raises the rate on every future request; attaching the guardrail to one more application raises the volume.

## Safety spend, kept proportional

ZopNight's Bedrock provider discovers guardrails on the 6-hour cycle and tracks evaluation cost from Cost Explorer or CUR 2.0 with usage trends. Guardrails are not a resource anyone should race to switch off, and the posture reflects that: the value is proportionality, keeping the inspection surcharge visible next to the inference it protects. A guardrail whose evaluation cost rivals its application's model cost signals a configuration worth examining. The usual cause is maximal policies applied uniformly where the risk profile varies by endpoint.

## Where inspection over-applies

The common excesses: every policy type enabled on every guardrail because the console made it one checklist, billing grounding checks on traffic with no retrieval to ground; internal and experimental workloads inheriting the production guardrail, paying safety rates to inspect developers talking to test prompts; and duplicated evaluation, where both an application-level and a team-level guardrail inspect the same exchange twice.

## Guardrail configuration review

The Bedrock console's Guardrails view lists each guardrail with its policies and versions. The review pairs each attached application with the policies it actually needs. Per-endpoint tailoring is where the surcharge and the risk profile get back into proportion.