Which model for what
Every assistant offers several models, and the names change every few months. The three classes behind them stay – and with them the rule for when the thorough model is worth it.
Boaz Lichtenstein
Your assistant’s model picker shows four or five entries today, and in six months they will have different names. Memorise the names and you relearn them every few months. Know the classes behind them and all you have to do is sort.
Three classes instead of twenty names
Fast. Small models answer in a second or two and barely touch your quota. They are good at rephrasing, summarising, translating and sorting – anywhere the task consists of one step.
All-rounder. The default in almost every assistant. Writes well, understands longer documents, occasionally takes one step too few on nested tasks.
Thorough. These models visibly “think” for longer: they break the task down, work out intermediate steps, discard an approach and check the result against the question. That takes anywhere from ten seconds to several minutes and costs a multiple of the quota.
What longer thinking actually changes
A thorough model does not know more than the fast one of the same generation. Its knowledge ends on the same day, and a figure it does not know it still does not know after two minutes. What changes is the care: fewer skipped steps, fewer contradictions between beginning and end, better plans.
That gives you the mapping:
The automatic mode – and when to overrule it
Many assistants now decide for themselves how much effort a question gets. That is usually right, with two typical misses: a briefly worded but hard question is taken too lightly. And a long but simple request triggers the expensive mode for no reason.
Overrule the automatic mode in two cases. First, when a mistake would be expensive – a number that goes out, a legal assessment, a decision about money. Second, when the first answer was shallow: a longer follow-up rarely helps then; the same question one class higher does.
The rule of thumb
Start with the default. Switch up when the task has several steps that depend on each other or an undetected error would have consequences. Switch down when you need a lot of the same. And record once what switching up costs you – in waiting time and in your plan’s quota. How the billing behind it works is covered in Understanding AI costs.
What no model replaces is a good brief. A thorough model with a vague brief thinks for a long time about the wrong question – the four parts from Prompts that hold up apply in every class.