The choice you have been ignoring
Most tools let you pick a model. Bigger models think better and cost more. Smaller ones are
faster and cheaper.
Most people pick one and never change. That leaves real money and real time unclaimed.
A reasonable way to choose
Use a smaller, faster model for mechanical work. Renaming things across files. Writing
tests from cases you supplied. Formatting. Simple, well-defined changes where the answer is
obvious and only the typing is slow.
Use a stronger model for thinking. Debugging something you do not understand. Designing
how a feature should work. Anything where being subtly wrong is expensive — which, as you
learned in Module 4, is most things involving rules and dates.
Use a stronger model when you cannot check the work easily. If your tests will catch a
mistake in ten seconds, a cheap model is fine, because being wrong is cheap. Where you
cannot check easily, pay for being right the first time.
That last rule is the useful one, and it connects the whole course: how well you can
verify decides how much you should pay. Good tests buy you cheaper models.
Cost habits that actually help
Start fresh sessions often. Lesson 1.5: the whole conversation is resent every message.
This is usually the biggest saving available.
Do not load files you do not need. Lesson 6.4. Cheaper, faster, better answers.
Use helpers for searching. Lesson 6.5. Thirty files read in a helper’s context, three
lines returned to yours.
Watch for loops. An agent stuck retrying the same fix burns money fast. Lesson 4.6 said
stop after three attempts. That advice has a price attached.
On speed
Faster is not always better. A model that produces something in five seconds that you spend
ten minutes checking has not helped.
Judge by the time to working, checked, committed code. That is often better with the
slower, more careful option.
Try this before the next lesson
- Find where your tool shows usage. Which session cost the most? Why?
- Do one mechanical task with a small model and one with a large one. Any difference in the
result? - Which of your recent tasks could you not have checked easily? Those deserve the better
model.