Model routing replaces single-model deployment with three orchestration patterns — Single, Cascade, and Critique — now standardised by GitHub Copilot Fusion (GA 28 Oct 2026). Fusion acts as a runtime orchestration engine across VS Code, CLI, JetBrains, and the Copilot app, selecting models per prompt from a pool including Claude Opus 5.5, Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol, billing each at standard rates with a 30% discount through 30 Nov 2026. Artificial Analysis reported Fusion Max at $5.59/task versus $11.42 for Claude Code on Opus 5.5 Max (~51% cheaper, within two quality points, 40% faster), though this remains a vendor-published benchmark. The Cascade pattern — cheap draft, quality gate, optional escalation — is the cost-optimisation path; its economics hinge entirely on the judge. A weak gate escalates everything (no savings) or accepts garbage (no quality). The article provides a Python cascade skeleton using regex complexity signals (multi-file, concurrency, debug keywords, >120 tokens), mocked cheap/flagship clients at $0.15/M and $3/M tokens, and a toy judge checking draft length and code presence. With 1,000 daily requests (2k tokens, 80% pass rate), all-flagship costs $6.00/day; cascade costs $1.95/day (67% saving), assuming a free judge. An LLM judge adds token cost, shifting break-even; deterministic judges (tests, lint, schema) are cheapest. Escalate pre-draft on strong complexity signals; use Single for routine edits, Critique for high-stakes refactors. Open questions: router drift as model prices shift, LLM judge calibration standards, Fusion discount expiry, and local open-weight drafters for data sovereignty.