It's Not Which AI You Use, It's Which Level

A friend tried ChatGPT once, asked it something, got a confidently wrong answer, and decided the whole thing was overhyped. Maybe you know that friend. Maybe, once, you were that friend.

Here’s the thing worth knowing before you write AI off: they probably didn’t meet a bad model. They met a good model on its lowest setting.

The question almost everyone asks is “Which one is best, ChatGPT, Claude, or Gemini?” It’s a fair question, and it has answers. But it’s not the question that decides whether AI feels like a party trick or a genuine help. That one is quieter, and almost nobody asks it: at what level am I actually using this?

Think of it like a ladder. Same tool, three rungs, and the view changes completely as you climb.

The reframing, in one sentence

There’s a growing pile of evidence that the system around a model, whether it can take its time to reason, search the web, remember you, use tools, touch your files, moves real-world results more than swapping one brand-name model for another.

To put a number on it: researchers at MIT’s FutureTech group looked at what they call “scaffolding,” the software wrapper that turns a raw model into a capable assistant, and found it explains more of the difference in price and performance than the choice of model does. In the strongest cases, giving a model a better setup was worth about two years of model progress.

In plain terms: a modest model with good tools, web access, and memory will often beat a flagship model stuck alone in a bare chat box. That gap between the two is the whole story of this article. So let’s climb the ladder.

Level 1: The free chat box

This is where almost everyone meets AI, and it’s the weakest rung, which is a shame, because it’s also the one people judge everything by.

You open a free account on the website or app, you type, it types back. It’s genuinely capable, and for a lot of things it’s all you need. But three quiet limits shape the experience.

It thinks less. The newest models can “reason,” meaning they take a moment to work through a hard problem step by step before answering, but on free plans that ability is trimmed down. As of mid-2026 the free tiers have finally gained a light thinking mode (a smaller reasoning model on ChatGPT, standard thinking modes on Gemini, limited extended thinking on Claude), so the old line “free means no thinking at all” isn’t quite true anymore. But the deep, patient reasoning still lives behind the paywall.

It runs out. Free plans cap how much you can use them, often a short burst of messages, then a cooldown of a few hours. Right when you’re getting somewhere, it asks you to wait.

It knows less about now. Web search and memory are thinner or absent, so answers lean on what the model absorbed during training. That’s exactly the recipe for the confident-but-wrong answers we wrote about in an earlier piece on hallucinations, and exactly why the free box feels like it “makes stuff up.”

The free level isn’t bad. It’s just the ground floor. The mistake is standing on the ground floor and concluding the building has no upstairs.

A fair thing to say: for rewriting an email, explaining a concept, or kicking around ideas, free is completely fine. The trouble starts when someone uses it for fact-heavy or multi-step work, gets a shaky result, and quietly decides AI isn’t for them.

Level 2: The paid chat box

For somewhere around €20 a month, the ladder’s second rung looks like the same chat box. It isn’t. What you’re really buying isn’t “more messages,” it’s a set of new abilities that switch on.

It can properly think. That deep, step-by-step reasoning mode is now yours on demand. Ask it a genuinely hard question and it takes its time instead of blurting the first plausible thing.

It can look things up. Reliable web search, with sources you can click. Time-sensitive questions stop being educated guesses and start being answers you can check.

It remembers you. Across conversations, it holds on to your context, what you’re working on, how you like things done, so you stop re-explaining yourself every morning. (Claude actually brought memory to its free tier back in March 2026; on the others it’s more of a paid-tier strength.)

It works with your stuff. You can upload files for it to read, set up dedicated project spaces with your own instructions, and connect things like your Drive or email so it works from your real material instead of thin air.

What this means for you: this is the rung where AI stops being a toy and becomes a tool you’d actually trust with a real work email, a piece of research, or planning your week. If you’ve only ever used the free box and found it underwhelming, this is the single upgrade most likely to change your mind.

Level 3: The desktop agent

Here’s the rung the free-only user will never even see, and right now it’s where the biggest jump happens.

On the first two levels, the AI talks. It answers, explains, drafts. On this one, it works. Through the desktop app, Claude’s “Cowork” or ChatGPT’s “Work” mode, the assistant stops being a voice in a box and becomes something closer to a colleague who can actually do the task.

What changes is that it gets hands. It can read and write real files in a folder you point it at. It can build a spreadsheet with working formulas, not a rough table you have to clean up, but a proper Excel file that calculates. It can assemble a slide deck, open a browser and click through websites, and even run tasks on a schedule while you’re asleep. It breaks a big job into steps and works through them.

A concrete example, straight from the tools’ own documentation: drop three months of receipts into a folder and ask it to sort out your expenses. On the free box, it would explain how to organize receipts. On the paid box, you’d upload them one by one and it would read them and hand you a table to copy out. On the desktop agent, you point it at the folder, walk away to make coffee, and come back to a finished Excel file, receipts read, amounts totalled, saved back where you can find it.

That’s the leap. Not a better answer. A finished thing.

A few honest caveats, because this rung is powerful:

It’s not free of mistakes. An agent taking real actions can get real things wrong, which is why these tools are built around approval steps. They show you a plan, and they ask before doing anything consequential. Claude’s Cowork, for instance, won’t delete a file without your explicit yes.

There’s a new kind of risk. Because an agent can act on your data and browse the live web, a booby-trapped web page can try to hijack it with hidden instructions, a trick called “prompt injection.” Both companies treat this as a headline safety problem, but it’s worth knowing the risk exists before you hand over the keys.

And it costs more. This kind of multi-step work burns through usage limits faster than a simple chat, so it’s for the jobs that are worth it, not for asking what year a film came out.

For most beginners, this rung is optional today. But it’s plainly where things are heading, and it’s the clearest reason the “which level” question matters more than the “which brand” one.

And a rung beyond

There’s one more level, and we’ll only wave at it: wiring a model into your own custom setup through its programming interface, the world of automations and business systems. It’s where companies build AI directly into their workflows, and it’s exactly where that MIT finding bites hardest, since the custom scaffolding can matter as much as the model. It’s a developer’s playground, not a beginner’s. Just know the ladder keeps going.

When the model does still matter

None of this means the brand on the logo is irrelevant, let’s be fair. For the genuinely hard frontier stuff, advanced mathematics, competitive-grade coding, expert reasoning at the edge, the top models really do pull ahead of the rest, and the choice matters.

And the picture keeps shifting: free tiers are slowly gaining abilities that used to be paid-only, so the gap at the bottom is narrowing, not widening.

But for the things most of us actually do with AI most days, the pattern holds. A bigger model badge won’t rescue you if you’re stuck on the wrong rung, and a very ordinary model on a higher rung, good tools, live web, real memory, the ability to act, will quietly run circles around a flagship trapped in a plain chat window.

Find your level

So, where are you right now?

If you’re on a free plan, typing into a box and sometimes rolling your eyes at the answers, you’re on Level 1, and honestly, you’ve been judging a whole ladder by its bottom step.

If you pay for a plan and use thinking mode, web search, and file uploads, you’re on Level 2, and you’re already getting most of what makes AI genuinely useful.

If you’ve handed a folder to a desktop app and watched it build something, you’re on Level 3, and you already know why the other two feel slow.

Here’s one small experiment for each rung, to try this week:

On Level 1, before you write AI off, turn on the “thinking” or “reasoning” option for one hard question and watch the answer get better.

On Level 2, make a project, drop in a few of your own files, and ask it something that only makes sense with your material in front of it.

On Level 3, give a desktop agent one boring, finite job, organize a messy folder, or turn a pile of notes into a clean document, and see how it feels to get back a finished file instead of instructions.

The best AI isn’t a secret model that only insiders know about. Most of the time, it’s the one you already have, used one rung higher.

Sources

• Just a Wrapper? How Much Do Scaffolds Matter? (MIT FutureTech)
• Get started with Claude Cowork (Anthropic Help Center)
• Introducing ChatGPT agent (OpenAI)
• What is the Pro plan? (Anthropic Help Center)
• Gemini plans and Deep Think (Google) ��������������������������������������������������������������������