Which model, once you have decided it is AI
Deciding a part of the job needs judgement leaves you one more choice: which model does it. It is tempting to want that answered with a name, but the name is the part that moves. Providers add models, retire them, reprice them and shift what each one is good at, several times a year. What stays put is the shape of the trade, so it is worth learning the shape rather than the lineup.
There are broadly three tiers. The trade between them is the same wherever you look. A big model that reasons hardest and costs most. A middle one that gets close for much less. A small one that is fast and cheap and will not carry difficult thinking.
| What matters most | Reach for | The trade |
|---|---|---|
| The hardest reasoning, plus long autonomous runs where being right matters more than the bill | The top tier | It costs the most per token and it is the slowest to answer. On routine work it deliberates well past the point of usefulness. |
| Most work. The balance point between quality and what you pay for it | The middle tier | Close to the top on the great majority of tasks, for a fraction of the price. This is the default, not the compromise. |
| Speed and volume on short, well-scoped work: classifying, extracting, tidying | The small fast tier | It will not carry the hard reasoning, so the saving turns into a wrong answer if you send it there. |
The instinct is to reach for the top tier on everything, on the grounds that the best answer is worth having. Usually it is not. The middle tier handles the great majority of real work at a fraction of the cost. The money saved there is what pays for the top tier on the handful of things that genuinely need it. Reaching for the biggest by default is not caution, it is just not having chosen.
One more thing, because it catches people out: the tier is not the only dial. Most models now take a separate setting for how hard to work on a given task. Turning that down is often a better saving than dropping a tier, because it keeps the model that actually understands the problem. Which tier is which today, what each one is called and what it costs are all live details, so read them at the source rather than from a page like this one: choosing a model in the Anthropic documentation has the current lineup. The tiers are durable. The names are not.
Part of When to use AI and when not to.