kasrin.com

1 min read

The model barely matters

The industry line on AI coding is simple: pick the right model and your problems are solved. If not today’s model, tomorrow’s will do it.

50 hours and a production healthcare app later, I learned that the model barely matters.

I ran 12+ open-weight models through the same harness on the same codebase. The differences were marginal. I spent €15 on tokens total. That €15 wasn’t an achievement — it was proof that the model was never the bottleneck.

External data backs this up:

The pattern is clear: model size and price don’t predict outcomes for development tasks. Security doesn’t improve with bigger models. Task completion doesn’t require frontier models.

What did matter was the environment I built around the model — the harness, the task scoping, and the documentation structure.

So stop agonizing over model selection. Pick any competent open-weight model, avoid lock-in to proprietary offerings, and move on. The model isn’t where your time goes — and it shouldn’t be where your attention goes either.

Having freed yourself from the model question, the next thing to look at is the harness. That’s the next piece.

Published on August 2026


Other Writings →