Describe the feature
When using Amazon Bedrock, some models do not support the CountTokens API (for example, cohere.embed-v4:0). For these models, applications still need a reliable way to obtain the exact input/output token counts after an InvokeModel request for purposes such as usage metering, billing, quotas, and analytics.
Currently, token counts may be available as HTTP response headers (e.g. x-amzn-bedrock-input-token-count and x-amzn-bedrock-output-token-count), but accessing these headers is not straightforward or consistent across SDKs and versions.
Feature Request
Please expose token usage returned by the Bedrock service through a stable, documented SDK interface.
For example, one of the following would be helpful:
response.$metadata.tokenUsage = {
inputTokens: number;
outputTokens: number;
};or
response.tokenUsage = {
inputTokens: number;
outputTokens: number;
};The exact shape is less important than having a documented and supported way to access this information without relying on raw HTTP headers.
If exposing these values is not possible because the Bedrock service does not guarantee these headers, clarification in the documentation would also be appreciated.
Use Case
Applications integrating Amazon Bedrock often need exact per-request token usage to:
Track tenant usage
Attribute costs
Enforce usage quotas
Build usage analytics
Today, this is difficult for models that do not support CountTokens, as applications must rely on transport-specific behavior instead of a supported SDK interface.
Proposed Solution
If the Bedrock service already returns token usage information, expose it through a typed field on the InvokeModel response or within $metadata.
If the service does not currently return this information in a guaranteed way, consider exposing it as part of the service response so SDKs can surface it consistently across languages.
Other Information
CountTokens is not supported for all Bedrock models (for example, cohere.embed-v4:0).
This feature would provide a consistent SDK experience across models and eliminate the need for applications to depend on undocumented HTTP response headers or custom middleware.
Acknowledgements
SDK version used
@aws-sdk/client-bedrock-runtime@3.1079.0
Environment details (OS name and version, etc.)
- OS: macOS Tahoe (Apple Silicon M4) - Runtime: Node.js v24 - Language: TypeScript
Describe the feature
When using Amazon Bedrock, some models do not support the
CountTokensAPI (for example,cohere.embed-v4:0). For these models, applications still need a reliable way to obtain the exact input/output token counts after anInvokeModelrequest for purposes such as usage metering, billing, quotas, and analytics.Currently, token counts may be available as HTTP response headers (e.g.
x-amzn-bedrock-input-token-countandx-amzn-bedrock-output-token-count), but accessing these headers is not straightforward or consistent across SDKs and versions.Feature Request
Please expose token usage returned by the Bedrock service through a stable, documented SDK interface.
For example, one of the following would be helpful:
or
The exact shape is less important than having a documented and supported way to access this information without relying on raw HTTP headers.
If exposing these values is not possible because the Bedrock service does not guarantee these headers, clarification in the documentation would also be appreciated.
Use Case
Applications integrating Amazon Bedrock often need exact per-request token usage to:
Track tenant usage
Attribute costs
Enforce usage quotas
Build usage analytics
Today, this is difficult for models that do not support CountTokens, as applications must rely on transport-specific behavior instead of a supported SDK interface.
Proposed Solution
If the Bedrock service already returns token usage information, expose it through a typed field on the InvokeModel response or within $metadata.
If the service does not currently return this information in a guaranteed way, consider exposing it as part of the service response so SDKs can surface it consistently across languages.
Other Information
CountTokens is not supported for all Bedrock models (for example,
cohere.embed-v4:0).This feature would provide a consistent SDK experience across models and eliminate the need for applications to depend on undocumented HTTP response headers or custom middleware.
Acknowledgements
SDK version used
@aws-sdk/client-bedrock-runtime@3.1079.0
Environment details (OS name and version, etc.)