Everyone's asking "which AI model should we use?"

Everyone's asking "which AI model should we use?"

Most of the time, that's the wrong question.

A study out on arXiv this week put a number on it: small, cheap models handled routine business tasks at about 90% of a frontier model's performance. For roughly 4% of the cost. Not by finding a smarter model. By building a smarter setup around the model.

In plain English: the "harness" is everything wrapped around the AI: the instructions you give it, the tools you connect it to, the way you break a job into steps. The researchers found that most of what makes a task hard doesn't live in the model at all. It lives in that wrapper. Move the difficulty into the wrapper, and a 4-cent model can do work you were paying a dollar for.

Here's why that matters if you run a business (and aren't a machine learning engineer):

The model is the part you don't control. It's someone else's product (priced how they want, changed when they want). That wrapper is the part you own. That's the part you can actually shape, and it happens to be the cheap part.

I've been building things on the web long enough to have seen this movie before. The winners are rarely the ones with the fanciest single component. They're the ones who got the boring plumbing right.

Two honest caveats:

• This works best for repetitive, well-bounded jobs (processing invoices, sorting tickets, pulling data). For genuinely open-ended judgment, you still want the stronger model.
• "Cheaper" only counts if the setup is good. A small model with a lazy wrapper just fails faster and cheaper.

So before you upgrade to the priciest model on the menu, try a different question: have we actually invested in how we use the one we've already got?

That's usually where the money is hiding.