There is a new best model on the market. But take care, it should be reserved for your "most complex" tasks, because it will burn more tokens than the smaller models you should be using for the "simple" ones. And (bonus!) you can now choose how hard it thinks. So now you really can pick the right model for every task and set the effort to match. Powerful model for the hard work, cheap model for everything else, tune as you go. Do it right and you will avoid the kind of eye-watering bills that not even Uber and Microsoft can keep a lid on, having adopted AI just a little too successfully.

Or you could ignore all that.

I am not going to argue that matching the model to the task is wrong in principle. It is sound, and when you are setting your organisation's AI budget for FY27 and beyond, it will matter a great deal. But for you, right now, there are more important things to focus on. So use the best model available to you, leave it on its sensible default, and stop touching the settings. If your tool offers to choose the model for you, let it.

Start with what it even means for a task to be complex. I could not reliably work it out. I tried, and went down a rabbit hole: complex compared to what, and to whom? It turns out the things these tools find hard are not the things I find hard. And if I am honest, even if I could sort the complex from the simple, I would never remember to change the model, let alone change it back.

The distinction barely matters anyway. I pay a fixed monthly fee that caps my usage, and on the rare occasions I hit the cap I take it as a sign to do something that does not need AI, like have a coffee. You are probably in the same boat. For now at least, this is one of the most heavily subsidised ways to use AI there is - the major labs are still selling these subscriptions for well below what heavy use actually costs them.

The real reason to leave the settings alone is not the bill. It is that the effort you would spend learning to route work between models is effort not going into the one skill that actually decides whether AI is any good in your hands. And unlike model selection, that skill does not expire when the next model ships. It is your judgement: your ability to decide what to use AI for, and to evaluate what it gives you back and know whether it is right.

That second part is the one that matters, because it is the one the tool cannot supply. The models are now good enough that the output almost always looks right. It is fluent, confident, well-structured, and sometimes wrong in ways you will only catch if you already know what good looks like. That recognition is the whole game, and it is the thing you bring, not the thing the model brings.

Here is the part most people get wrong. Judgement like this does not build by using AI more. The research points the other way: the people who use these tools most heavily, and who know the most about them, tend to be the most confident and the least accurate about their own work. Volume builds the feeling of competence without the substance. You can spend a year using AI every day and come out more sure of yourself and no better at telling good from plausible.

What builds the judgement is a small habit that costs you almost nothing. Form your own view first. Before you read the AI's answer, decide what a good one looks like - what it should contain, where it might go wrong, what you would expect a sharp colleague to say. Then read what the tool produced and notice the gap. Sometimes it catches something you missed. Sometimes you catch something it got confidently wrong.

Either way, you just did the thing that turns use into capability: you checked your judgement against an outcome, instead of adding another hour of use.

There is a stronger version of this, and it is the one I would teach first. When something matters, do not ask the tool to do it for you. Ask it to argue against you. Give it your view and tell it to make the strongest case that you are wrong, then judge which of its objections actually land. You get a better answer, and you get the rep that builds the judgement, in a single move. The model becomes your most useful critic instead of your most flattering assistant.

This is also why staying put helps. Each model behaves slightly differently, with its own habits and its own blind spots, and you only learn to catch those by working with one long enough to feel them. A new top-tier model has arrived from one of the major labs in twelve of the last fifty-two weeks. The twelfth arrived while I was writing this.

If you let it, the switching never stops, and you stay a permanent beginner. Better to know one tool well enough to catch it when it is wrong than three too shallowly to notice. Most of the major tools have now started making the model choice for you, behind the scenes - which should tell you something about how much it was ever worth doing by hand.

And do not wait. The flat-fee arrangement that lets you ignore the model choice is already ending - the labs have begun moving their newest models out of subscriptions and onto usage-based pricing. When the meter goes on, routing work between models will become a decision with a bill attached, and the people who make it well will be the ones who can already judge what the tools give back. You cannot pick the right model for a task if you cannot tell whether its output is any good.

So this week, do one thing. Use the best model available to you, or let the tool pick if it offers, and decide you are done choosing. Then the next time it matters, form your own view first, and make it argue against you. That is the work that compounds.

PS for the legal readers, and anyone whose client work runs under strict confidentiality terms: the new top-tier model arrives with a condition its predecessors did not have. Prompts and outputs are retained for thirty days for safety monitoring - on every platform, including where your organisation holds a zero-data-retention agreement. Nothing changes for the rest of the lineup, and nothing changes at all unless your firm turns the new model on. But if your engagement terms depend on zero retention, put this past whoever owns your AI vendor arrangements before anyone upgrades.