Inference Cost Baseline and Margin Guardrail (Onboarding)

Inference Cost Baseline and Margin Guardrail (Onboarding)

1. 01

   Inventory every model call the client solution will make before quoting a retainer

   List each distinct task (summarization, classification, retrieval, agent orchestration) and the expected monthly call volume. A single unlisted high-volume task can swing infrastructure cost by four figures.

2. 02

   Price the workload against at least two providers at current published rates

   Anthropic's Claude Haiku 5.5 landed at $0.10 per million input tokens with a 1M context window, so a workload priced on an older frontier model may be 5x to 10x more expensive than necessary.

3. 03

   Separate fixed platform cost from variable token cost in the client-facing estimate

   Hosting on Render, Fly, or Vercel carries a predictable floor; token spend scales with usage. Blending them hides which line item breaks the margin when volume doubles.

4. 04

   Set a monthly token budget ceiling and name the person who approves overages

   Route spend through a gateway such as Helicone, Portkey, or TrueFoundry so the ceiling is enforced in code rather than discovered on an invoice.

5. 05

   Model the cost at 3x projected volume and confirm the retainer still clears target margin

   Client pilots routinely grow faster than forecast. If 3x volume erases margin, reprice the scope or move the workload to a smaller model before signing.

6. 06

   Document which tasks can degrade to a cheaper model without visible quality loss

   Classification and routing rarely need a frontier model. Liquid AI's open-weight d1-3B and d1-omni-600M decision models exist specifically for tasks that produce no output tokens.

7. 07

   Record the baseline in the client agreement as a reviewed assumption, not a fixed promise

   Provider pricing moves. A clause that ties the retainer to a quarterly cost review protects the agency when a vendor changes rates mid-contract.