Which AI model should you use. Small, big or reasoning.

Open Claude or ChatGPT and you meet a menu of names: Opus, Sonnet, Instant, Pro. Behind every menu are just two dials: size (how big the brain is) and thinking (how long it plans before answering). Thinking happens in what the model writes, because it first writes hidden notes to itself before answering.

Ask a model what 47 times 38 is. Answering in one leap often lands wrong. With notes first, it works out 47 × 40 = 1,880, then 47 × 2 = 94, and reads the answer, 1,786, straight off its own notes. A reasoning model is trained to write useful notes, with steps, checks and fixes.

What each effort level actually costs

Every hidden note is one more full pass, so thinking means buying more work per question:

• Low: ~500 tokens, ~5 seconds, a quick answer
• Medium: ~3,000 tokens, a bit more waiting, a solid answer
• High: ~10,000 tokens, closer to a minute, a checked answer
• Max: tens of thousands of tokens, minutes of waiting, a double-checked answer

You pay for thinking even though it's hidden. In one sample receipt, 3,000 hidden tokens cost six cents, the biggest line on the bill, because everything the model writes counts as output.

Matching the dial to the job

Ask a top model on max effort to fix typos in a caption and you get the same answer a fast model gives in a second. Ask what the risks in a lease are, and a fast model on low effort misses the auto-renewal clause, while a capable model on high effort flags it and the rent increase.

A simple rule: the best model is the smallest one that does the job well. A newer model on low effort often beats an older model on max. Before choosing an AI product, ask which model handles your requests, whether it thinks harder on hard questions, what that costs, and whether it was tested on tasks like yours.

For a deeper look, see reasoning models vs chat models or our step-by-step guide to choosing an AI model. For enterprise AI strategy, explore Silverberry's Fractional Chief AI Officer.