-
Split Thinking From Typing
Use your strongest model for the part where mistakes are costly: understanding the codebase, choosing architecture, finding edge cases, and writing the build plan. Once the plan is clear, hand implementation to a cheaper model that is good enough at producing code, edits, docs, or repetitive assets.
-
Make The Spec The Contract
Before switching models, ask the stronger model for a detailed implementation spec. Include files to touch, expected behavior, constraints, test cases, migration notes, and what not to change. The cheaper model should not be asked to rethink the problem, it should be asked to follow the spec.
-
Bring The Expensive Model Back For Review
After the cheap model implements the work, send the diff or PR back to the stronger model for review. This uses the frontier model where it has high leverage: spotting design flaws, missed requirements, and risky shortcuts. Feed only the review comments back to the cheaper model for fixes.
-
Lower The Dial For Rote Work
Do not use top-tier models with high reasoning for deployments, formatting, simple copy edits, CSV cleanup, or mechanical refactors. Pick a lower model tier and lower reasoning effort for tasks where the desired output is obvious. The habit matters because default settings quietly burn budget.
-
Watch Output Tokens
Code generation, document drafting, and long rewrites create lots of output tokens, which are usually the expensive part. Planning often reads more than it writes, so premium-model cost is more contained there. The routing rule is simple: pay for judgment, discount the volume.
Why it matters
Small businesses do not need enterprise tooling to get the benefit of model routing. Even manual copy-paste between tools can stretch a fixed AI budget and keep better models available for decisions that matter. This is especially useful for solo builders, where every wasted premium run competes with product work, customer work, or cash.