Token Cost Intelligence — OpenClaw Optimization Framework

Token costs at scale are not a billing problem — they are a routing problem. Production API models are 10-50x more expensive than local inference for work that does not require them. Classification, formatting, summarization, extraction — these do not need Claude Sonnet. They need a fast, cheap model that returns the right answer. This SKILL.md is the routing framework that makes that happen automatically.

---

⚡ What's Inside

• Three-tier compute model Task classification across production API (complex reasoning, novel generation), mid-tier inference (summarization, extraction, formatting), and local/free inference (classification, routing decisions, low-stakes logic)

• Task taxonomy 40+ task types pre-classified by tier with the reasoning behind each classification. Stops you from re-litigating these decisions every time

• Cost instrumentation How to track token spend by model, by task type, and by day. The instrumentation spec that makes invisible costs visible before they become a problem

• Routing decision tree The five-question evaluation sequence that classifies any task in under two seconds. Plugs into existing agent logic without architectural changes

• 8-10x cost reduction framework The specific configuration changes that produced an 8-10x monthly cost reduction in a live deployment. Not estimates. Measured outcomes.

---

🏭 Proven in Production

This framework came from a $15/day API spend that dropped to under $1/day after systematic task routing. The task taxonomy was built by auditing six weeks of actual token logs — every task type, every model call, every place where a $0.01 local inference call was happening as a $0.15 API call.

---

🆕 What's New in v1.1

• Local inference routing for Ollama/Vulkan backends added
• Daily cost report spec included with alerting threshold configuration

---

✅ Core Capabilities

• ✅ Three-tier compute model — production API, mid-tier, and local inference routing
• ✅ 40+ task types pre-classified — no re-analysis required for common workloads
• ✅ Cost instrumentation spec — daily spend tracking by model, task type, and role
• ✅ Routing decision tree — five questions, any task classified in under two seconds
• ✅ Local inference integration — Ollama/Vulkan routing for zero-cost classification tasks
• ✅ 8-10x cost reduction framework — the exact configuration that produced measured results