Amazon Bedrock Guardrails now supports Standard tier for contextual grounding checks

Amazon Bedrock Guardrails has launched a Standard tier for contextual grounding checks, designed to detect hallucinations by verifying whether model responses align with reference sources and user queries. The new tier replaces the Classic tier’s strict grounding approach, which was primarily suited for summarization, paraphrasing, and question answering. The Standard tier now supports responses that include reasoning or conversational content, offering more consistent scoring for differently worded but semantically equivalent answers.

The Standard tier enhances language support and aligns with the Standard tier for content filters and denied topics, providing more comprehensive coverage. Users can configure guardrails through the Amazon Bedrock console or API, selecting the tier for each guardrail. Existing guardrails default to the Classic tier until manually updated, with no change in pricing between tiers.

To enable the Standard tier, guardrails must use cross-Region inference with a guardrail profile. This requirement ensures consistent performance across supported AWS Regions, which include multiple locations in the U.S., Europe, and Asia Pacific. The feature is now available in 11 specified regions, expanding accessibility for developers building generative AI applications.

The update allows developers to apply contextual grounding checks to a broader range of applications, including conversational interfaces and reasoning tasks over source material. Amazon provides documentation and a product page for further details on contextual grounding checks and safeguard tiers. The Standard tier aims to reduce hallucinations while maintaining flexibility for different use cases.