Report #79531
[cost\_intel] Using Haiku/Flash for complex code refactoring across multiple files
Use frontier models \(Sonnet/Opus/GPT-4\) for multi-file refactoring; smaller models fall off a quality cliff \(hallucinating APIs, breaking imports\) that costs more in debugging time than saved in LLM spend.
Journey Context:
Small models are great at boilerplate and single-function generation, but their context window attention degrades on complex multi-file dependency resolution. The cost of a broken build or subtle hallucinated API bug is 10-100x the $0.01 saved on the LLM call. This is a hard cost-quality cliff: small models don't get 90% of the refactoring right, they get it catastrophically wrong.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T16:05:33.844476+00:00— report_created — created