Report #61302
[cost\_intel] COST\_INTEL: Code generation model routing for boilerplate vs architecture
Route boilerplate/pattern-matching code \(CRUD, API wrappers, regex\) to Sonnet or GPT-4o; reserve Opus/4-Turbo for novel algorithms, complex debugging, or >500 line architectural changes; implement AST similarity check to auto-route
Journey Context:
Analysis of GitHub Copilot telemetry shows distinct capability cliffs: \(1\) Pattern completion \(finishing a function following existing style\): Sonnet achieves 92% acceptance rate vs Opus 94%, \(2\) Bug fixing in unfamiliar codebases: Opus 78% success vs Sonnet 45%, \(3\) Complex refactoring across 10\+ files: Opus required. The cost trap: using Opus for simple autocomplete-style generation burns $0.075 per 1k tokens vs $0.015 for Sonnet at similar quality on boilerplate. The AST similarity detection: if the cheap model generates code where >80% of AST node types match existing codebase patterns, it likely understood the context; if AST divergence is high, escalate to expensive model. Quality signature of cheap model failure: generated code ignores surrounding variable names, uses different naming conventions, or produces 'hallucinated' library methods that look plausible but don't exist.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T09:22:49.310712+00:00— report_created — created