🎧 Listen to this article
Two groups are describing different planets
Last week Matt Shumer, an AI CEO and investor, published an essay about building entire software applications with AI. "I'll tell the AI: here's what it should do, here's roughly what it should look like. Figure out the user flow, the design, all of it," he wrote. "It writes tens of thousands of lines of code." The essay went viral. One group nodded, because they'd done the same thing. The other called it delusional, because their experience with AI was asking ChatGPT for a recipe and getting a mediocre one.
Both groups are describing their actual experience accurately. Both are wrong about each other. A paywall has split the AI user base into two populations using fundamentally different technologies, and the word "AI" now refers to both.
Only 3% have used the real thing
Menlo Ventures published the number in June 2025: 3% of AI users are paid subscribers. Matt Murphy, a partner at the firm who led investments in Anthropic, told CNN in February 2026 that he expects the ratio to shift. But today, 97% of people forming opinions about AI are using free-tier tools that have been deliberately capped.
The free tier of ChatGPT, Claude, and Gemini limits token counts, shortens context windows, runs older models, and strips agentic capabilities entirely. Someone who tried free ChatGPT to plan a vacation and got generic results isn't wrong to call AI overhyped. For them, it is. They used a product intentionally handicapped to create an upgrade incentive.
Emily DeJeu, a professor at Carnegie Mellon who teaches AI in business, told CNN it would be "misguided" to judge AI capabilities based on free tools. The blunter version: the most consequential technology debate of this decade is being conducted by people who've never used the technology they're arguing about.
The canyon between free and paid
Free-tier Claude writes paragraphs about supply chain management. Paid-tier Claude with extended thinking analyzes a 200-page contract, identifies liability clauses, cross-references case law, and produces a prioritized risk assessment. These aren't different degrees of the same capability. They are different capabilities.
Free-tier ChatGPT suggests workout routines. GPT-5.3 Codex builds, tests, and deploys full-stack applications from natural language. Shumer primarily uses Codex. His critics have never touched it.
Google's Gemini 3.1 Pro ships with the consumer app, but the actual capability lives in Antigravity, the agentic development platform available through enterprise and developer tiers. The consumer app gives you a chatbot. Antigravity gives you a platform for autonomous agents. Same model. Different product.
The canyon exists by design. Free tiers drive adoption: downloads, active users, the metrics that justify valuations. Paid tiers drive revenue. The gap drives conversion. If free were too good, nobody would pay. If too bad, nobody would try. The calibration is deliberate: impressive enough to create interest, limited enough to create frustration, just frustrating enough to sell upgrades.
When a pricing strategy becomes an information crisis
The 3%/97% split would be trivia if AI were a consumer gadget. It matters because AI is becoming infrastructure, and infrastructure decisions are being made by people on both sides of the canyon who don't know the other side exists.
Corporate executives who've seen free-tier demos underinvest because the demos are unimpressive. They watch a chatbot produce generic text and conclude the technology isn't ready. Their companies fall behind competitors who invested based on paid-tier capabilities. By the time the laggards recognize the gap, the leaders have compounding advantages in automation and cost structure.
Policymakers who form AI opinions from chatbot experiences write regulations calibrated for the wrong risk level. They either overregulate from science fiction anxieties or underregulate from "it can't even write a good email" dismissiveness. Neither calibration matches the production AI systems making financial decisions and generating code right now.
Knowledge workers who haven't used paid AI underestimate the threat to their roles. They see a tool that makes spelling mistakes and conclude their jobs are safe. Their employer is piloting agents on premium tiers that handle 40% of the workflow those employees perform. The adaptation window is narrow.
The strongest case for the skeptics
Shumer's credibility carries a 2024 asterisk. He apologized that year for exaggerating model performance, calling it the "biggest mistake" of his "professional life." His claims about AI building applications are almost certainly true in the narrow sense: the AI wrote the code. But he's vague about which model, what complexity the app involved, and how much time went to reviewing and fixing the output.
A 2024 Retool developer survey found 67% of developers using AI code generation spent significant time debugging AI-generated code. The productivity gain was real but smaller than the marketing. Paid-tier AI is genuinely more capable than free-tier AI. Whether it's capable enough to justify the revolutionary rhetoric is a fair question.
The rebuttal is temporal. The skeptics' evidence comes from 2024 tools. AI capabilities compound with each generation. ARC-AGI-2 scores doubled between Gemini 3 and 3.1. The Arena leaderboard climbs steadily. Codex handles tasks in February 2026 that were impossible in February 2025. Skepticism calibrated to 2024 capabilities is already outdated, and the pace of change means it becomes more outdated every quarter. The free-tier skeptics aren't wrong about what they've seen. They're wrong about what exists beyond what they've seen.
What happens while the world catches up
Companies led by executives who've crossed the paywall pull ahead of companies led by executives who haven't. The performance gap compounds quarterly. This shows up in earnings calls where AI-forward companies report automation savings and AI-skeptical companies report margin pressure from competitors who did invest.
The adaptation window for knowledge workers narrows. Professionals who learn to work alongside paid AI tools position themselves for the labor market of 2028. Those who don't are optimizing for a market that's disappearing. This isn't abstract. The 40% workflow coverage that enterprise agents handle today was 15% in 2024 and near zero in 2023.
The paywall isn't a pricing strategy. It's the most consequential information barrier in technology. Three percent can see through it. Ninety-seven percent see only their own reflection. The technology works. The information about what it does hasn't reached the people who need it most. Until it does, two groups will keep describing different planets and calling each other liars.
Sources: Menlo Ventures (June 2025), Matt Murphy/CNN (February 2026), BNN Bloomberg (February 21, 2026), Emily DeJeu/Carnegie Mellon, Matt Shumer (February 2026), Retool Developer Survey (2024)
