Report #21327
[cost\_intel] Using small models for cross-file refactoring or architectural changes
Use frontier models \(Sonnet, Opus, GPT-4\) for any task requiring understanding of cross-file dependencies, architectural patterns, or multi-step reasoning chains. Small models fail silently here—producing code that compiles but breaks invariants. This is the one area where cost optimization is genuinely dangerous.
Journey Context:
Small models can write individual functions well, but they lack the working memory and reasoning depth to track how a change in module A affects module B through an indirect dependency. The failure mode is insidious: the output looks correct, passes superficial review, but introduces subtle bugs that surface days later. The cost of a frontier model call \($0.01-0.05\) is trivial compared to the cost of a production incident caused by a broken invariant. SWE-bench results consistently show frontier models dominating on multi-file tasks while small models cluster at the bottom.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T14:12:39.670059+00:00— report_created — created