Which AI model should you use- Part 2 - How AI thinks

Why does the same AI cost a fraction of a cent one minute and twenty cents the next? Two dials: model size and thinking effort. Together they decide whether you get a quick answer, a checked answer, or expensive overkill.

How thinking actually works

A thinking model writes hidden notes to itself before answering. Each note word is one more full pass through the network — so thinking adds care, while size adds judgment. A small model with notes can even rescue a tricky sentence like "she sat on the bank" by reasoning from context to "river."

Every answer burns three kinds of tokens:

• Input — what the model reads
• Thinking — hidden notes you don't see but still pay for
• Output — the visible answer

Matching effort to the task

Rough shape of the dial:

1. Low — ~500 tokens, ~5 seconds, quick answer.
2. Medium — ~3,000 tokens, ~15 seconds, solid answer.
3. High — ~10,000 tokens, ~45 seconds, checked answer.
4. Max — up to ~30,000 tokens and minutes, double-checked.

Costs multiply across both dials: small-and-light can run a twentieth of a cent, while large-and-thinking can hit twenty cents — roughly 400x apart. Fixing a headline typo on max effort is overkill; running an expense report on the fastest model with low effort can miss duplicate receipts a capable model on high effort catches.

Before picking an AI product, ask three questions: which model handles your requests, does it think harder on hard questions and at what cost, and was it tested on tasks like yours? The rule of thumb: pick the smallest model and the least thinking that does the job well — and check you're on the latest model before turning effort up.

For a deeper look at the pricing side, see hidden reasoning tokens and what you actually pay for, or the step-by-step guide to choosing effort level. Teams building this into real workflows can talk to our AI consulting team.