Amazon Bedrock

The models of Amazon Bedrock.

Features

Provides llm, text-embedding, rerank models in Dify.
Includes predefined llm models such as openai, mistral, ai21.
Includes predefined rerank models such as amazon.rerank-v1, cohere.rerank-v3-5.
Includes predefined text embedding models such as cohere.embed-multilingual-v3, cohere.embed-english-v3, amazon.titan-embed-text-v2.
Supports predefined model and customizable model configuration.

Setup

Install this plugin from the Dify Marketplace.
Get the required credentials from Amazon Bedrock.
Add the credentials in the plugin settings.
Save the configuration.

CloudWatch Request Metadata (Optional)

When enabled, attaches and as on Bedrock Converse API calls. This allows attributing individual invocations to a Dify app when querying CloudWatch model invocation logs.

Purpose: Log attribution for per-request analysis. does not appear in Cost Explorer or on the bill.
Requirements: Model invocation logging must be enabled on the AWS account.
Scope: Only applies to Converse API calls (not other API types).
Default: Disabled by default, opt-in via the credential setting.
How to enable: Set the "Enable CloudWatch request metadata" option to "Enabled" in the provider or model credentials.

Temporary Credentials

The plugin supports AWS temporary credentials from SSO/SAML authentication (e.g., or ):

Access Key authentication: optionally provide an AWS Session Token for temporary credentials. The plugin will use the session token alongside your access key and secret key.
IAM Role authentication: credentials are automatically refreshed on every invocation. The plugin creates a fresh boto3 session for each IAM Role authentication call, which picks up the latest credentials from disk (via , , IMDS, etc.). This proactive approach eliminates errors and makes credential rotation transparent without requiring plugin daemon restarts.

Usage

Select Amazon Bedrock as the model provider in Dify, choose an available model, and use it in applications, agents, or workflows.

Claude 5 Models (Opus 5.5 / Sonnet 5.5 / Fable 5.1 / Opus 5 / Sonnet 5 / Fable 5)

The "Anthropic Claude 5" entry provides Claude Opus 5.5 (), Claude Sonnet 5.5 (), Claude Fable 5.1 (), Claude Opus 5 (), Claude Sonnet 5 (), and Claude Fable 5 (). These models differ from earlier Claude generations:

Inference profiles only. Invocable exclusively via cross-region inference profiles — there is no on-demand bare-ID invocation. Geographic profile coverage varies per model: Opus 5.5, Opus 5 and Sonnet 5 offer (also serves Canada regions), , and profiles; Sonnet 5.5 offers and ; Fable 5.1 and Fable 5 offer only. Opus 5.5 also has a profile, which the plugin does not select yet (use from Japan regions). works from virtually all commercial regions. The plugin resolves this from your Cross-Region Inference selection; when your region has no geo profile, choose .
Adaptive thinking is on by default. Thinking depth is controlled by the Effort parameter (////) instead of a thinking budget. Sampling parameters (temperature / top_p / top_k) are not configurable on these models and are not exposed.
Refusals & fallback. Requests can be declined by safety classifiers (; materially more frequent on Fable 5 and Fable 5.1). With Refusal Fallback enabled (default), the plugin automatically retries the identical request with Claude 4.8 Opus. Mid-stream refusals that occur after output has been produced cannot fall back and raise an error instead.
Fable 5 prerequisite: data retention opt-in. Your AWS account must set data retention mode to via the Bedrock Data Retention API before invoking Fable 5 (no console UI at launch). See the Fable 5 model card. Fable 5.1 needs the same opt-in (Fable 5.1 model card). The plugin never changes account settings.
Pricing note. Displayed prices are the Global cross-region rates (Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, Fable 5.1 $10/$50, Opus 5 $5/$25, Sonnet 5 $3/$15, Fable 5 $10/$50 per 1M tokens). Geo/in-region invocation is ~10% higher. Prompt-cache read/write rates are not representable in Dify's cost tracking — refer to your AWS bill.

Japan data residency ( cross-region inference)

Selecting for Cross-Region Inference uses the geographic profile, which routes only to and — inference never leaves Japan. It requires the configured region to be one of those two, and the model to have a profile (Sonnet 4.5 / 4.6, Haiku 4.5, Opus 4.7 / 4.8 and Nova Lite V2 at the time of writing). Otherwise the call fails with an explanatory error instead of silently routing elsewhere.

is not equivalent: from Tokyo it resolves to , which also routes to Korea, India, Singapore and Australia — and no profile exists for 4.5-generation models or later. The pricing note above applies here too: displayed prices are the Global cross-region rates, and geo/in-region invocation is ~10% higher.

Privacy

This plugin sends the inputs required by the selected operation to the upstream service. See PRIVACY.md for details.