# Engineering FinOps Playbook: 65/35 Frontier-to-Fast Model Cascade & Prompt Caching

> Deterministic FinOps playbook routing 65% of sub-turn discovery, linting, and unit-test prompts to DeepSeek-V3.2 / Haiku while pinning 35% of architectural diffs to Claude Sonnet 4.6 with Anthropic ephemeral KV caching.

- Canonical: https://pulse-forge.io/playbooks/litellm-finops-cost-router-playbook
- Download Free Starter CLAUDE.md ($0): https://pulse-forge.io/api/launch-kit/free-claude-md?repo=BerriAI/litellm
- 1-Click $4.99 Drop-In Config Pack: https://pulse-forge.io/buy/micro_pass?repo=BerriAI/litellm&source=md_playbook
- 1-Click $9.99 Complete 9-File .ZIP: https://pulse-forge.io/buy/repo_pass_999?repo=BerriAI/litellm&source=md_playbook_999
- 1-Click $49 Verified Leaderboard Boost: https://pulse-forge.io/buy/instant_boost?repo=BerriAI/litellm&source=md_playbook_49
- Reference Repo: `BerriAI/litellm` (-62.4% token waste)