Choosing a model for the job
In one line: ask what it costs you to be wrong, and whether the text is allowed to leave your machine. Those two answers pick the model.
The decision, in ten seconds
1. What does a wrong answer cost me?
Nothing much — I will spot it immediately. → the small fast model. Tidying notes, drafting an email, summarising, renaming, sorting, "what is the word for…".
Something — I might not spot it. → the mid or large model. Analysis, a decision you will act on, code that goes somewhere real, anything where a plausible wrong answer would slip past you.
2. Does this text need to stay on my machine?
Yes. → a local open-weight model, and nothing else will do. Client material, anything under NDA, medical, financial, personal.
No. → whichever hosted model suits question 1.
That is it. Everything else is refinement.
What actually changes between the tiers
Not "how much it knows" — they all know roughly the same things. What changes is how much work it does before answering.
A small model answers in one pass. A large reasoning model effectively thinks before it speaks: considers approaches, checks itself, notices when the first idea does not work. That is why it is slower and dearer, and it is why the difference shows up on hard problems and almost vanishes on easy ones.
So the rule is not "use the best you can afford". It is "use the smallest one that will not be wrong in a way you would miss."
A rough map
| What you are doing | What suits it | | --- | --- | | Tidy, summarise, rewrite, extract | small / fast | | Draft something you will edit | small or mid | | Answer a factual question you will check | mid, with search if the answer changes over time | | Work through a real problem | large / reasoning | | Code you are going to ship | large, or a coding tool like Claude Code | | Anything confidential | local, whatever it costs you in quality |
Two traps
Reaching for the biggest model by reflex. It is slower, it costs more, and on routine work it is no better. If you would spot a mistake in half a second, you do not need a model that thinks for thirty.
Asking a model about anything recent without giving it a way to look. It has a training cut-off. Ask about this week's news and you will get a confident answer built from last year. If the answer could have changed since it was trained, either give it the current facts yourself or use a tool that searches.
What AIOS does about this
AIOS picks for you. It works out what a request needs before spending anything: most of what you type is handled with no model call at all — a capture, a completion, a reschedule — and only a genuine question goes to one. When it does, it picks the cheapest model that fits and moves up only if that one cannot cope.
You do not have to think about any of this to use AIOS. This lesson is so that you know what it is doing, and so you can make the same call yourself when you are using something else.
Try this
Take a task you did with AI this week and ask: would a smaller model have been fine? Most of the time the answer is yes, and noticing that is most of what "being good at this" amounts to.
---
Next: Asking for what you actually want · All lessons