The story of the ugly cat already introduced it: running on fumes. A poor drawing happened because, on a tight day, three things were dialled down at once. Those three things aren't an anecdote — they're the basic frame of all AI use: three dials that together determine how good an answer gets and what it costs. Understand them once, and you'll read model names, subscription tiers and a chatbot's "form of the day" differently.
Dial 1: model class
AI providers ship their models in size classes, like engines. At Anthropic, for instance, the current range runs from the small, fast Haiku through the mid-range Sonnet up to the top-tier Opus and Fable models. The developer price list (for people building the models into their own applications) makes the spread tangible — as of July 2026, per the official models overview: Haiku costs 1 US dollar per million processed text units ("tokens"), Sonnet 3, Opus 5, Fable 10 — with generated text costing five times as much in each class. Between the smallest and the largest model lies a factor of ten.
The point isn't the dollar figure — as a chat user you never pay it directly. The point is: intelligence is billed by the token, and bigger models cost more per token. Every subscription, every free tier, every "unlimited" is a quota on this resource. Small models aren't "bad", by the way — for summarising, rephrasing and everyday questions they're fast and entirely sufficient. Class matters where things get hard: layered problems, long context, conceptual work.
Dial 2: thinking budget
Modern models can work internally before answering — playing through intermediate steps, discarding approaches, double-checking. This "thinking" consumes additional tokens, invisibly to you, and it can be dosed: from "answer immediately" to "take your time". Current models increasingly decide this themselves based on task difficulty; providers and tools additionally offer levels to cap or extend the thinking depth.
For you this means: the same question to the same model can yield answers of different quality — depending on how much thinking budget the application grants at that moment. A tightly capped top-tier model can lose to a mid-range model that's allowed to think.
Dial 3: number of attempts
The underrated dial. First answers are rarely the best — good AI work happens in loops: look at the result, verify it (or have it verified), improve. For code the loop is called "run the tests", for the cat it was "look at the screenshot", for prose it's "read it critically". Every round costs tokens again — but it's often the difference between usable and good. Note that just saying "again, but better" doesn't turn this dial: that's repetition without new feedback.
The quiet dial-down
Now for the interplay hardly anyone accounts for: you are not the only one operating these dials. The provider turns them too — sometimes silently. When your quota runs low or demand peaks, some chat interfaces switch unannounced to a smaller model or trim the thinking budget. Answers then get quietly worse. You can't see the low flame in a mediocre paragraph — you can see it in a cat.
So if a chatbot that was brilliant yesterday seems scattered today: that need not be your imagination. Check which model is selected, whether your quota is running out, whether the interface has an "auto" mode choosing on your behalf.
Setting the dials deliberately
As an everyday rule of thumb: simple things small and fast — summaries, phrasing help and everyday questions don't need a flagship model. Hard things big and with room to think — concepts, analysis, tricky debugging deserve the best available class. Important things with several attempts — and with genuine checking between rounds, ideally by you. It only gets expensive when all three dials sit at maximum permanently; used deliberately, the spread is exactly what makes AI affordable.
/compact — the essentials, if context is running low:
Three dials together determine how good an AI answer gets and what it costs: the model class, the thinking budget for internal reasoning, and the number of attempts — everything is billed in tokens. Providers turn these dials too, sometimes silently: when a quota runs low or demand peaks, some chat interfaces switch unannounced to a smaller model or trim the thinking budget, and answers get quietly worse. As a rule of thumb: simple things small and fast, hard things big and with room to think, important things with several attempts and genuine checking between rounds.