View a markdown version of this page

Monitor bedrock-mantle inference using CloudWatch metrics - Amazon Bedrock
Inference metricsToken metricsDimensionsChoosing the right granularity levelDifferences from bedrock-runtime metricsViewing bedrock-mantle metrics

Monitor bedrock-mantle inference using CloudWatch metrics

The Amazon Bedrock OpenAI-compatible endpoint (bedrock-mantle.region.api.aws) publishes customer-facing metrics to Amazon CloudWatch under the AWS/BedrockMantle namespace. Use these metrics to monitor inference volume, token consumption, and error rates across your AWS account, projects, and models.

If your application calls inference through bedrock-runtime.region.amazonaws.com using Converse, ConverseStream, InvokeModel, or InvokeModelWithResponseStream, see Amazon Bedrock runtime metrics instead. bedrock-mantle metrics cover only inference issued through the Responses API, Chat Completions API, and Anthropic Messages API on the bedrock-mantle endpoint.

bedrock-mantle publishes metrics at four levels of granularity. Each level uses a different combination of CloudWatch dimensions, described in Dimensions. All metrics are scoped to your AWS account.

Inference metrics

Inference metrics
Metric name Unit Description
Inferences Count Total number of completed inference requests across the Responses, Chat Completions, and Messages APIs. Published at Account, Project, Model, and Project+Model levels.
InferenceClientErrors Count Number of inference requests that failed with a client-side (4xx) error. Published at Account, Project, Model, and Project+Model levels.

Token metrics

Token metrics
Metric name Unit Description
TotalInputTokens Count Aggregate billable input tokens processed within the publish interval. Published at Account, Project, and Model levels. Use for sums and rate calculations.
TotalOutputTokens Count Aggregate billable output tokens generated within the publish interval. Published at Account, Project, and Model levels.
InputTokens Count Per-inference billable input tokens. Each completed inference emits one datum. Published at the Project+Model level only. Use this when you need percentile statistics (p50/p90/p99) of per-request token counts.
OutputTokens Count Per-inference billable output tokens. Each completed inference emits one datum. Published at the Project+Model level only.

Dimensions

Dimensions
Dimension name Values Applies to
Project The project ID associated with the inference request. Project, Project+Model
Model The model identifier (for example, anthropic.claude-opus-4-7). Model, Project+Model

Account-level metrics are scoped to your AWS account and have no additional dimensions. Each level of granularity emits independently, so a single inference contributes to all four levels when both the project and the model can be resolved.

Choosing the right granularity level

  • Account level – overall usage, error rate, and aggregate token volume. Suitable for high-level dashboards and account-wide alarms. Not suitable for cost analysis, because pricing varies by model.

  • Project level – per-project rollups for chargeback and team-level dashboards.

  • Model level – per-model usage and error rates, suitable for migrating dashboards built against the existing bedrock-runtime ModelId dimension.

  • Project+Model level – primary level for cost analysis, and percentile latency and token analysis. Use this when you need both project attribution and model attribution on the same datum.

Differences from bedrock-runtime metrics

  • Separate namespace. bedrock-mantle metrics publish to AWS/BedrockMantle. Existing dashboards and alarms built against AWS/Bedrock will not pick up bedrock-mantle traffic.

  • Naming. bedrock-mantle uses Inferences rather than Invocations, TotalInputTokens and TotalOutputTokens rather than InputTokenCount and OutputTokenCount, and InferenceClientErrors rather than InvocationClientErrors.

  • Project dimension. bedrock-mantle metrics carry a Project dimension that bedrock-runtime metrics do not, enabling per-project cost attribution.

  • Cross-region inference. bedrock-mantle is in-region only. Metrics are emitted in the Region that handled the request and are not aggregated across Regions the way cross-Region inference (CRIS) traffic is on bedrock-runtime.

  • Latency metrics. InvocationLatency and TimeToFirstToken equivalents are not yet published by bedrock-mantle.

Viewing bedrock-mantle metrics

To view bedrock-mantle metrics in the CloudWatch console:

  1. Open the CloudWatch console.

  2. In the navigation pane, choose Metrics, All metrics.

  3. Choose the AWS/BedrockMantle namespace.

  4. Select a granularity level by choosing the dimension set that matches your query (for example, Project, Model for project-and-model breakdowns).

You must have appropriate CloudWatch permissions to read bedrock-mantle metrics. For more information, see Authentication and access control for Amazon CloudWatch in the Amazon CloudWatch User Guide.