How I Choose My LLM Models for Token Efficiency

Early on, I used to struggle with deciding what model to choose to maximize my token-usage efficiency.

But the first time I heard about an approach for this, it made so much sense that I haven’t looked back.

The rule is simple: treat agents as you would  treat human developers.

Use the more capable (and expensive) ones for high-level tasks, and keep the less capable (and cheaper) ones for low-level tasks that require less reasoning powers.

In practical terms, how does this look in my workflow? As follows:

I use expensive models for the kind of thing I would do with a Principal-level developer: discuss new features, prepare implementation plans, discuss trade-offs, review PRs from other models, and from time to time, even ask for its opinion regarding product-related decisions I’m about to make.

As for the cheaper models, I use them as a very capable Junior engineer: when plans are done and all details are defined, I ask them to implement everything I discussed with the more capable model.

I’ve read extensive articles about this, but it boils down to a very simple rule:

Plan with the expensive agent, build with the cheap one.

And I believe that the beauty of this approach is in its simplicity: it is definitely tempting to go on a never-ending cycle of research to determine exactly which model is more appropriate for each task, but in my opinion, all of that takes time away from the actually most important task:

To build.