Report #90877
[cost\_intel] Using Haiku/Flash for multi-file refactoring or complex algorithm implementation
Route multi-file refactors and complex logic generation to frontier models \(Sonnet/GPT-4\). Smaller models produce syntactically valid but logically flawed code.
Journey Context:
Smaller models are great at boilerplate and single-function generation. But they fall off a cliff on tasks requiring >5 steps of logical reasoning or cross-file dependency tracking. The quality degradation signature is hallucinated imports and wrong state mutations. The human debugging cost exceeds the inference savings.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T11:07:59.491025+00:00— report_created — created