OpenAI spend per feature — the leak is often the SKU, not the model

Chat, summary, and copilot share a model and still lose money in different places. A summary that runs on every ticket will outspend a chat that a few seats use. Reprice the SKU, not the model card.

Send x-cs-feature on the same request as x-cs-tenant and x-cs-key. The customer P&L splits OpenAI spend per feature. Point the client at https://api.costslice.com/v1.

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.costslice.com/v1",
    api_key=os.environ["OPENAI_API_KEY"],
    default_headers={
        "x-cs-key": os.environ["COSTSLICE_KEY"],
        "x-cs-tenant": tenant_id,       # opaque id, never an email
        "x-cs-feature": "support-copilot",
        "x-cs-env": "prod",
    },
)

Put OPENAI_API_KEY and COSTSLICE_KEY in your product’s environment variables. We never store the OpenAI API key.

Tenant tags: OpenAI cost per tenant. Workflows: n8n OpenAI cost. OpenAI cost P&L.

Worked example: on the sample, clinical-summary is the SKU that sinks acme-health. Chat on another customer can still look healthy on the same stack. Reprice the feature, not the model. See the sample P&L.