Watch two people argue about AI sometime. One of them says it’s a toy. It makes things up, it forgets what you told it ten minutes ago, the code it writes doesn’t run. The other one says it has quietly taken over half his working day, and he’d sooner give up his truck.
Here’s the part almost nobody says out loud: they’re probably both right. They’re just not describing the same machine.
The first guy is on a free plan. He gets a capped number of messages in a five hour window, and when he hits the wall, the app invites him to pull out a credit card. That’s not me being cynical, that’s OpenAI’s own help page: free users get the flagship model “only a limited number of times within a five hour window,” and the published cure for hitting a rate limit is upgrading your plan. The second guy is paying $200 a month for Claude Max, which Anthropic sells with a number right in the name: 20x. Twenty times the usage of the $20 plan, on the heaviest models, all day.
Same word, “AI.” Two completely different machines.
For one strange year, everybody had the same thing
Rewind to the end of 2022. ChatGPT shows up, it’s free, and there is no better version. A college kid, a plumber, and a Fortune 500 strategy team all typed into the same box and got the same brain back. For about a year, the most capable AI on earth had a price of zero, and the research from that window is genuinely remarkable.
A study published in Science gave real writing tasks to 444 working professionals and handed half of them ChatGPT. Everyone got faster and everyone’s work got better, fine, you’d expect that. The interesting part is who gained the most: the weaker writers. Before ChatGPT, being a weak writer on the first task pretty much guaranteed you’d be a weak writer on the next one. With ChatGPT, half of that gap between the weak writers and the strong ones simply disappeared. The researchers said it themselves: initial inequalities were half erased.
The same shape showed up everywhere people looked. A study of thousands of customer support agents found the new and low skilled folks got 34% better with an AI assistant while the veterans barely moved. At BCG, one of the biggest consulting firms in the world, 758 consultants ran realistic work tasks with GPT-4. The weakest half of the consultants improved 43%. The strongest half improved 17%. The gap between the best and the rest shrank, and the Harvard team behind the study said it plainly: AI levels performance differences.
That is the leveling everyone felt in 2023. It was real, it was measured, and it had an asterisk nobody read: everyone in those studies was handed the same model.
Then the menu showed up
Today nobody hands you the model. You pick it off a menu, and the menu runs from $0 to $200 a month per person. Free tiers cap your messages and throttle the good features, the $20 tiers raise the ceiling, and the $100 to $200 tiers sell usage in multiples, plus the heavy models and the tools that hang off them. The capability wall is still there for everybody. The difference is that now it moves when you pay.
So who pays? Almost nobody, it turns out. Menlo Ventures did the math on roughly 1.8 billion people using consumer AI worldwide: only about 3 in every 100 pay anything at all. Even ChatGPT only convinces about 5 in 100 of its weekly users to subscribe.
Now hold that against the people who use AI for a living. In a survey of over 7,200 software developers, the question was “which of these models do you or your company actually pay for.” Only about 1 in 7 said none. Read those two numbers together. The general public overwhelmingly experiences AI through capped free tiers. The professionals whose paycheck depends on it overwhelmingly pay.
I went looking for a study that splits “AI is amazing” against “AI is useless” by subscription tier, and I couldn’t find one. Nobody seems to have run it yet. So I’ll mark this next sentence as my read, not a citation: when one group of people lives on quota-capped free models and another lives on frontier models with the limits paid off, the argument about whether AI is any good stops being an argument about AI. It’s two groups reviewing two different products that happen to share a name.
The field is tilting, and someone measured it
The institutions have started measuring the tilt. An OECD report from June 2025 tracked business AI adoption across wealthy countries and found the leaders pulling away from everyone else: the spread between the most-adopting and least-adopting countries widened from 14 points in 2021 to 24 points in 2024. In their words, the growth is “more driven by leaders escaping the pack than laggards catching up.”
Inside that same report is the number that should bother a small business owner. In 2024, 39% of large companies were using AI. For small businesses it was 12%. Big companies are adopting AI at more than three times the rate of small ones, and that’s a wider big-versus-small gap than any technology wave the OECD has tracked. Social media never split big and small businesses like this. Neither did the cloud. AI is.
Microsoft Research found the same pattern down at the neighborhood level: in the first six months of ChatGPT, interest clustered in educated, well-off counties, and education was the single strongest predictor of who even knew about it. Their warning: left alone, these early differences “show a potential to reinforce existing spatial and socioeconomic divides.”
A tool that measurably lifted the bottom hardest is being adopted fastest by the top. Both things are true at once, and the second one is winning.
The same line runs through software
I see this divide most clearly in my own trade. The cheap end of AI assisted software is “vibe coding”: describe an app to a $20 builder, accept whatever comes out, ship it. And the volume of working-looking output is honestly impressive, which is the trap.
Veracode’s 2025 security research found that when AI writes code, it introduces a known security flaw almost half the time. GitClear studied 211 million lines of code changes and watched copy-pasted code pile up while the cleanup work that keeps software healthy fell by more than half. And AI will hand you the flawed line with exactly the same confidence as the good one. It is brilliant, and it has no judgment about its own work.
The expensive end isn’t expensive because the typing is faster. It’s frontier models plus a person with judgment plus checks around everything the machine touches. The cost difference between those two builds is real. So is the difference between an app that takes payments and an app that leaks them.
What a small operator should actually do
Here’s the part that should be encouraging: the gap between big and small companies isn’t really a money gap. The scary $200 tier works out to about $6.60 a day, which is nothing next to what a business spends on labor. What the big companies are actually buying is experience. People who work with this stuff every day, who know which jobs AI is genuinely good at, which jobs it will quietly botch, and what kind of checks to build around it.
You don’t need to become that person. You run a business, and the last thing you need is a second job studying model tiers and usage quotas. The small businesses I see catching this wave aren’t the ones grinding away at it alone on a free plan. They’re the ones who brought in somebody who already lives on the strong side of the line, pointed them at one boring, repeatable job, and let a human with judgment decide where AI belongs in their business and where it doesn’t.
The leveling that those 2023 studies measured is still available to small businesses. It just stopped being the default, and these days it usually arrives through a partner instead of a login.
That’s the work we do at Allen App Tools, if you ever want to talk it through: a short call where we look at your business together and find where AI actually makes sense for you, and where it doesn’t.