Some coworkers and I have become those people. We now set aside an hour every week just to talk about what we built or learned with AI. We are such geeks that show and tell has become a recurring meeting, except now the toys can reason, call tools, and occasionally wander off and build way more than anyone even thought to ask for. Yesterday, one of the geeks asked a great question: “What do you guys do with models and reasoning? Are you choosing one and sticking with it, or do you find that you need something like Fable?”
I gave him a lousy answer. It wasn’t wrong. I started telling him what I do. Lately (and this will date this article!), I’ve been using Fable and GPT-5.6 Sol for heavy design work when I know I want as much intelligence as I can get. I usually leave reasoning on high, or its equivalent, because I’ve found that lowering it can burn more tokens in the end. The model gets something wrong. I explain it again, add guardrails, run it again, then fix whatever still went wrong. Lazy me would rather explain the problem once. I move to xhigh when I want that little extra mile of work. I want the model to double-check its reasoning, look at the answer again, and see whether it missed something. In my head, every step up the reasoning ladder feels something like a doubling. That isn’t scientific. It is just how I think about it so I don’t waste effort.
Then there are the ideas where I’m willing to wait an hour. Maybe I just had a breakthrough. Maybe I am thinking about a new project. I want the strongest mind I can get, and I want it to sit with the idea for as long as it can. That is when I reach for Fable or Sol with a crazy amount of reasoning.
Again, none of that was wrong. It answered what I do. It did not answer what he was really asking, which was how to think about any of this without first spending an unreasonable amount of time doing it. I kept thinking about his question after the meeting, and I think this is the answer I should have given him.
Choose the kind of mind. Then decide how long you want it to think.
Einstein and Shakespeare were both clearly brilliant, but not in the same way. Einstein gives us mathematical abstraction and physical reasoning. Shakespeare gives us language, character, ambiguity, and human motives. You wouldn’t choose between them by asking which one had more intelligence, as though intelligence were a tank of gas. You would ask what kind of problem was sitting in front of you.
Models are like that too. They have different weights, training, habits, and tendencies. They are all smart these days, including models you can use for free, but they are not the same kind of smart.
Reasoning effort changes how much time and work a model spends on the question. It does not turn one kind of mind into another one. This is where the question gets more interesting. Can a light or medium Einstein beat an xhigh Shakespeare? Of course. If the problem depends on seeing the right abstraction, more time may not give the other model the conceptual tool it was missing. It may just produce a longer walk down the wrong road. The reverse can also be true. A model with less raw capability may do better with more reasoning when the work is within its reach but needs patient checking, evidence gathering, or just the discipline to keep going. Sometimes the smarter model was simply too quick. Sometimes the slower answer is better. This is why treating models as one ranked list has never felt quite right to me.
I use models from several frontier vendors, and over time I have developed biases. I’ve even taught my agents to choose particular models for particular kinds of work. In my defense, I was recently hit by a car and have been unable to walk while recovering from surgery. It turns out that leaves a person with far too much time in the company of artificial intelligences for his own good. These are working impressions, not a leaderboard, and the models change too quickly for one anyway. I currently reach for OpenAI when I care about prose or want the model to think outside the box. Claude is neat and orderly. It tends to arrange the work well, although I often need to give it more of the frame. Grok is the rough kid. The bad boy. It can be useful when I want a devil’s advocate who may say the uncomfortable thing. Gemini used to feel like the artistic one with a bit of a hacker in it. I’m not entirely sure who Gemini is today. My researcher, maybe. I haven’t tried it in that role yet.
That’s the basic version, but even within a provider, the models exhibit different behaviors. And don’t get me started on the plethora of models outside these well-known names. Take something like Fable. The quality I noticed was relentless go. Give it a tiny seed and it becomes the Energizer Bunny. It keeps looking at the unfinished edges and driving toward something comprehensive. It was the first model where I felt that tendency so strongly. That has uses. It also has drawbacks. If I want a quick answer, I do not want Einstein, and I definitely do not want Einstein deciding that my quick question is the beginning of a new field of study.
Cost gets strange too. I’ve been amazed by what I can pull out of free models, even for really difficult work, including cybersecurity and agentic SOC applications. There are real bargains out there. When I don’t know which model or reasoning level to use for something new, I usually start by asking. I ask that model, and sometimes other models, what they know about it. I never trust the answer. Instead, I experiment. I give the models real work and see how they perform. Very unscientific, yes, but semantics have always felt a little unscientific to me. Our words are imprecise. It seems fitting that large language models are equally imprecise. From a business perspective I look for bargains. From a quality perspective I reach for the big guns. You generally get what you pay for, but the cheapest token and the cheapest finished result are not the same thing. A less expensive model with lower reasoning can cost more if I have to keep correcting it, adding instructions, narrowing the frame, and running the work again. High reasoning is partly a quality decision for me. It is also, honestly, the lazy choice.
So, back to his question. Do I choose one model and stick with it? No. I choose the kind of mind I want and how long I want it to think. Increasingly, I also think about what I expect it to do when the first obvious answer is not enough. This isn’t scientific, and it isn’t meant to be. It is just the answer I wish I had given yesterday, before I started listing model names and reasoning settings like the room already knew how I had arrived at them.
I still probably spend too much time thinking about this. At least I have a better answer for the next show and tell.
These views are my own and do not represent my employer. AI was only slightly harmed in the production of this article.