What a model costs, and how the trade-offs are made

The Beginner level gives you the selection criteria. Here, we look at what they actually cost, because as soon as you go beyond occasional use, price is what decides it.

You pay per token, in both directions

A subscription service hides the real billing from you. As soon as you go through an API, or automate anything at all, you pay per unit: what you send, and what comes back.

Output typically costs three to five times more than input. The direct and counter-intuitive consequence: asking for a short answer doesn't just save reading time, it cuts the bill. "Answer in three sentences" is a budgetary instruction as much as an editorial one.

Context is paid for again on every turn

This is the most expensive trap, and the least visible. In a conversation, the history is sent back in full with every message. By the twentieth exchange, you're paying twenty times for the beginning.

Hence a rule that has nothing aesthetic about it: one conversation per topic, then start clean. That's not tidying up, it's spending.

Caching

Most providers charge far less for a request prefix they've already seen. If you send the same long preamble every time, always place it in the same spot and at the top: it will be cached. Moving it by a single character cancels the benefit.

What the rankings are worth

Benchmarks measure specific tasks, often academic ones, and the gaps at the top come down to points no one perceives in practice. Two serious caveats: test sets eventually leak into training data, which inflates scores, and a model can be optimised for the test.

The only ranking that counts for you is the one you run yourself: take ten real cases from your work, run them through two models, compare. Half a day, and the result is true for you.

The trade-off rule

The most capable model for what is rare and high-stakes. The most economical for what runs at volume. Mixing the two in a single pipeline is almost always the right setting, and almost never the one people adopt spontaneously.