Use this checklist before choosing a model, agent loop, or automation route for a real task. Cost discipline is not token anxiety. It is the habit of matching the route to the value, risk, context, and review burden of the work.
The practical rule:
Use premium models when they change the economics of valuable work. Route commodity work down when scripts, templates, saved workflows, local models, or lower-cost models are good enough.
Step 1: Name The Work
Write the task in one sentence.
Task:
Expected output:
Who will use it:
Who reviews it:
What decision or workflow it affects:
If the task cannot be named clearly, do not spend more model time yet. Clarify the work first.
Step 2: Choose The Lowest Reliable Route
Start at the top and move down only when the cheaper route cannot meet the need.
- Existing system feature or deterministic code
- Template, saved workflow, or checklist
- Local or lower-cost model
- Mid-tier cloud model
- Premium model for high-value reasoning, synthesis, architecture, or ambiguity
- Human-only decision when accountability cannot be delegated
Selected route:
Why this route is enough:
Why cheaper routes are not enough, if applicable:
Step 3: Justify Premium Model Use
Use a premium model only when at least one of these is true:
- the decision is high leverage
- the task has ambiguity that cheaper routes handle poorly
- a bad output would be costly
- the result compresses meaningful expert time
- the work involves cross-source reasoning or architecture tradeoffs
- the output can still be verified
Premium route justification:
Verification owner:
Stop condition:
If the output cannot be verified, premium inference does not solve the problem. It only makes the uncertainty look better.
Step 4: Route Down Commodity Work
Route down when the work is mostly:
- extraction from structured inputs
- formatting or cleanup
- first-pass classification
- repeated summarization with stable source patterns
- draft generation where a cheaper route is good enough
- file inventory, counting, sorting, or other work better solved by code
Candidate cheaper route:
What quality check will prove it is enough:
When to escalate to a stronger route:
Step 5: Control Context And Retries
Large context and repeated retries can cost more than the task is worth.
Before running the task, decide:
- what source material is actually needed
- what can be summarized, filtered, or fetched precisely
- how many attempts are allowed
- what failure means stop instead of retry
- what reusable artifact should be saved if the task will repeat
Context limit:
Retry limit:
Reusable artifact to save:
Stop condition:
Step 6: Judge The Outcome Without Fake Precision
Do not invent ROI. Use practical evidence.
After review, answer:
- Was the output accepted, revised, or rejected?
- Did it reduce cycle time?
- Did it reduce missed follow-ups or manual rework?
- Did it improve decision quality?
- Did it create a reusable workflow?
- Did the route make sense compared with the value of the work?
Outcome:
Review effort:
Rework needed:
Reuse potential:
Decision for next time:
Route Decision Note
Copy and fill this block.
Task:
Selected route:
Reason:
Cheaper route considered:
Verification owner:
Context/retry limit:
Outcome evidence to capture:
Decision for next run:Common Mistakes
Watch for these:
- using a premium model because the task is annoying rather than valuable
- sending broad context when a smaller source packet would work
- retrying the same vague prompt instead of narrowing the task
- using a model for counting, sorting, or file inventory that code can do exactly
- treating low cost as permission to skip review
- treating high cost as proof the output is better
The best route is the cheapest route that can produce a reviewable, useful result for the work in front of you.