Cost Control

After traffic hits dxgate, Cost Control writes three distinct ledgers: Self-Hosted Throughput, Subscription Credits, and API USD. No artificial exchange rates. Unknown models are not priced.

Access: /ui → Cost Control, or via the management endpoint GET /debug/cost.

---

1. Three-Track Cost Model

Ledger Track	Account Pool	Metric Unit	Cost Attributes
Self-Hosted Compute	self-hosted	Concurrency / Token Throughput	$0 Marginal Cost (Tracks GPU load & queue latency)
Monthly Subscriptions	subscription	ChatGPT / Claude Credits	Fixed Monthly (Tracks period quota usage & reset dates)
Cloud Commercial API	api-key	Real USD (Tiered per Token)	Pay-As-You-Go (Calculates per-request USD ledger)

---

2. Subscription Configuration (Codex / ChatGPT / Claude)

Uses OAuth credentials with background auto-refresh. Quotas are recorded whenever usage is returned:

spec:
  ai:
    provider:
      openai:
        pool:
          - id: codex-sub-01
            type: subscription
            weight: 100
            credentialRef:
              name: oauth-codex-01
              key: token.json

Local Codex client configuration:

[model_providers.dxgate]
name = "dxgate"
base_url = "http://127.0.0.1:8080/v1"
wire_api = "chat"

[profiles.dxgate]
model = "gpt-5"
model_provider = "dxgate"

---

3. Pay-As-You-Go API Configuration

Inject API keys via Secret, clients connect via /v1:

kubectl -n dubbo-system create secret generic openai-secret \
  --from-literal=Authorization="$OPENAI_API_KEY"

spec:
  ai:
    provider:
      openai:
        pool:
          - id: openai-official
            type: api-key
            weight: 100
            credentialRef:
              name: openai-secret
              key: Authorization

---

4. Self-Hosted Compute Configuration

In-cluster vLLM / SGLang / Ollama clusters with $0 API fees, monitoring GPU throughput and concurrency limits:

spec:
  ai:
    provider:
      openai:
        pool:
          - id: vllm-deepseek-r1
            type: self-hosted
            weight: 100
            endpoint: http://vllm-service.ai-infra.svc:8000/v1
            maxConcurrency: 64

See LLM Services for full multi-account pooling and priority routing details.