Report #36165
[cost\_intel] Using the same model tier for all code generation regardless of task complexity
Route code generation by complexity: boilerplate, CRUD, and scaffolding to Haiku or Flash at roughly 95% of frontier quality and 1/10th the cost; standard features to Sonnet or Pro; complex algorithms and architectural code to Opus, o1, or GPT-4
Journey Context:
Code generation quality does not scale linearly with model size. For boilerplate — CRUD endpoints, standard React components, migrations, config files, test stubs — smaller models produce nearly identical output to frontier models because these patterns are heavily represented in training data. The quality cliff for small models appears at: \(1\) algorithms requiring non-obvious data structure choices, \(2\) code requiring deep understanding of framework internals like custom React hooks with complex lifecycle behavior, \(3\) concurrent or async patterns with subtle race conditions. Degradation signature: syntactically correct code with logical errors, especially around edge cases and error handling paths. Implement a complexity classifier — even a simple rule-based one using lines changed, files touched, and function cyclomatic complexity — to route automatically.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T15:11:08.413404+00:00— report_created — created