#Charging Back Token Usage

#Add the collection configuration

Cluster: the Cost Management agent cluster — the cluster where the AI Gateway runs and cost-agent is installed. The cost-agent component is deployed as the slark-agent workload, which the restart command below targets.

Create a ConfigMap that tells the Cost Management agent which Prometheus query to evaluate and how to map its labels onto Cost Management dimensions. The agent discovers configurations carrying the cpaas.io/slark.collection.config: "true" label in cpaas-system.

apiVersion: v1
kind: ConfigMap
metadata:
  name: slark-agent-aigateway-config
  namespace: cpaas-system
  labels:
    cpaas.io/slark.collection.config: "true"
data:
  config: |
    - kind: AIGateway
      category: AIGatewayToken
      item: AIGatewayTokenUsage
      period: Hourly
      usage:
        query: |
          sum by (user_namespace, "gen_ai.request.model") (
            increase({__name__="gen_ai.client.token.usage_token_sum"}[5m])
          )
        step: 5m
        mappers:
          name: gen_ai.request.model
          namespace: user_namespace
          cluster: ""
          project: user_namespace

Field reference:

• kind: the Cost Management collector kind, which also names the collector (keep it unique among collection configs). Pod is reserved for OpenCost-emitted pod metrics and Project for the platform's built-in CPU/Memory/Storage quota collectors. AI Gateway token usage uses a dedicated custom kind, AIGateway, matching the vGPU/NPU/pGPU custom-kind pattern — this is the value verified on the cluster to populate cost.usage with AIGatewayTokenUsage rows. After applying this configuration, confirm rows appear (see the cost.usage check in Troubleshooting) before moving on.
• category, item: identifiers used to link this collection configuration to its display/storage counterpart in the next step. Both values must match the corresponding fields in the display/storage configuration.
• period: the aggregation period. Use Hourly to bill by hour.
• usage.query: the PromQL query the agent evaluates every step. The platform stores the metric under its OpenTelemetry UTF-8 name with dots preserved, so select it with {__name__="gen_ai.client.token.usage_token_sum"} and reference the dotted gen_ai.request.model label with the quoted UTF-8 syntax. The header-mapped user_namespace label is the per-namespace billing key.
• usage.step: the query evaluation interval.
• usage.mappers: maps PromQL labels onto Cost Management's standard dimensions (name, namespace, cluster, project). Set cluster to an empty string so the agent fills in its own cluster identity automatically; setting it to a label name (such as cluster) only works if the source metric exposes that label, otherwise every row is dropped without a log entry.

After applying the YAML, restart the slark-agent workload (the cost-agent component) to reload the configuration:

kubectl -n cpaas-system delete pod \
  -l service_name=slark-agent --grace-period=0 --force
kubectl -n cpaas-system rollout status deploy/slark-agent
# if slark-agent is installed as a DaemonSet in your environment, use instead:
#   kubectl -n cpaas-system rollout status daemonset/slark-agent

#Add the display configuration

Cluster: the Cost Management server cluster — the cluster where cost-server is installed (typically the global control plane). The cost-server component is deployed as the slark-server workload, which the restart command below targets.

Create a ConfigMap that registers the billing item and its billing methods in the platform console. The server discovers configurations carrying the cpaas.io/slark.display.config: "true" label in kube-public.

apiVersion: v1
kind: ConfigMap
metadata:
  name: slark-display-config-for-aigateway
  namespace: kube-public
  labels:
    cpaas.io/slark.display.config: "true"
data:
  config: |
    - name: AIGatewayToken
      displayname:
        zh: "AI Token"
        en: "AI tokens"
      methods:
        - name: Usage
          displayname:
            zh: "使用量"
            en: "Token Usage"
          item: AIGatewayTokenUsage
          unit:
            zh: "tokens"
            en: "tokens"
          divisor: 1

Field reference:

• name: the billing item name shown in the cost model form. Must match the category value in the collection configuration.
• methods[].name: the billing method, listed under the billing item when adding a price.
• methods[].item: must match the item value in the collection configuration so the server can join the per-method price back to the usage rows.
• divisor: the unit conversion factor applied when displaying usage. Tokens are unitless, so set 1; for byte-sized items use 1073741824 to render as Gi-hours.

The following table summarizes the billing methods registered by this configuration:

After applying the YAML, restart the slark-server workload (the cost-server component) to reload the configuration:

kubectl -n cpaas-system delete pod \
  -l service_name=slark-server --grace-period=0 --force
kubectl -n cpaas-system rollout status deploy/slark-server

#Add a price to the cost model

Cluster: any cluster — operated from the platform console served by cost-api.

In the platform console, navigate to Administrator → Metering and Billing → Cost Model, then create or edit a cost model. The newly registered AI tokens billing item is now selectable in the price form.

1. Cost Model Name: any identifier, for example aml-cost-model.
2. Linked Clusters: select the clusters whose AI Gateway traffic this model should price. An empty selection saves successfully but matches no usage data and produces no bills.
3. Pricing rows:
   • Billing Item: AI tokens
   • Billing Method: Token Usage
   • Default Price: the per-token rate, in the platform's currency.
   • Price By Label (optional): per-model overrides, for example a higher rate for gen_ai.request.model="gpt-4o".
4. Save.