Supported Regions and models for Amazon Bedrock Guardrails
Amazon Bedrock Guardrails applies to foundation models that you invoke on the bedrock-runtime
endpoint through the InvokeModel,
InvokeModelWithResponseStream, Converse, and
ConverseStream operations. Within those requests, a guardrail can evaluate
both text and image content — for image content, see Block harmful images with content filters. The bedrock-mantle endpoint
doesn't support Amazon Bedrock Guardrails.
Most models in the Amazon Bedrock catalog support guardrails. The following models don't.
| Provider | Model | Model ID | Model type |
|---|---|---|---|
Amazon |
Titan Embeddings G1 - Text |
|
Text embeddings |
Amazon |
Titan Text Embeddings V2 |
|
Text embeddings |
Amazon |
Titan Multimodal Embeddings G1 |
|
Multimodal embeddings |
Amazon |
Amazon Nova Multimodal Embeddings |
|
Multimodal embeddings |
Amazon |
Rerank 1.0 |
|
Reranking |
Amazon |
Nova 2 Sonic |
|
Speech |
Amazon |
Nova 2.5 Sonic |
|
Speech |
Cohere |
Embed English |
|
Text embeddings |
Cohere |
Embed Multilingual |
|
Text embeddings |
Cohere |
Embed v4 |
|
Multimodal embeddings |
Cohere |
Rerank 3.5 |
|
Reranking |
Luma AI |
Ray2 |
|
Video generation |
OpenAI |
GPT 5.4 |
|
Text |
OpenAI |
GPT 5.5 |
|
Text |
TwelveLabs |
Marengo Embed v2.7 |
|
Multimodal embeddings |
TwelveLabs |
Marengo Embed 3.0 |
|
Multimodal embeddings |
TwelveLabs |
Pegasus v1.5 |
|
Video understanding |
Point-version model IDs follow their base model. For example,
amazon.titan-embed-text-v2:0:8k doesn't support guardrails because
amazon.titan-embed-text-v2:0 doesn't.
A request that references a guardrail with one of these models fails with the following error, even though the same request succeeds when you omit the guardrail parameters:
ValidationException: Guardrail is not supported with the chosen model
Note
Don't infer guardrail support from the model type. Models of the same type can differ — TwelveLabs Pegasus v1.2 supports guardrails while Pegasus v1.5 doesn't — and the Stability AI image generation models do support guardrails.
To evaluate content without invoking a model, use the ApplyGuardrail
operation. It applies a guardrail to content that you supply directly, so it isn't
tied to any model and works regardless of which models your application uses.
To confirm support for a specific model, go to models at a glance, choose the model, and check the Capabilities and Features section of its model card.
