ChatGPT Plus is a meter bolted to a machine nobody can read

In 2011, my first month in consulting, a manager handed me a six-digit copy code and explained, without any embarrassment, that every page I copied would be billed to the client as a line item. I stood at the machine with a ninety-page document and hesitated for the first time in my life over a photocopier, because a cost that had always been invisible had been made visible, and visibility turns out to be its own kind of tax. Then I watched what actually happened, which was nothing. Nobody at the firm counted pages; the copying ran at industrial volume for six more years; the code changed no behavior that anyone could name. What it changed was the accounting, and the accounting was the point. The client could be shown a number. I think about that code every time ChatGPT asks me whether I've considered upgrading, and I want to be precise about why.

ChatGPT Plus costs twenty dollars a month. What it buys is not intelligence. The free tier hands you the same model family with a smaller allowance. It buys ceiling: roughly 400 to 2,000 messages per five hours on the paid tier against about ten on free, plus Sora, Deep Research, Codex, Agent mode, and the rest of the 2026 bundle, per the pricing trackers that check these numbers against OpenAI's own pages. The product is the meter. And the interesting thing about this particular meter is that the company which installed it also decides, unilaterally and without publishing a number, how much it reads.

The cap is invented, and then it is sold back to you

Start with what OpenAI itself says about its limits, because it is less than you'd expect. The pricing page now lists everyday text chat as unlimited on every tier, footnoted with the requirement that usage be reasonable. The help article says subscriptions "may include usage limits such as message caps, especially during high demand," which "may vary based on system conditions." That is a meter with no dial. A power company that billed this way would be investigated. The pattern extends upward: in August 2026, OpenAI placed a five-hour usage limit on standard ChatGPT Business seats and then began selling a premium seat at $125 that removes it. Read that sequence carefully, because it is the whole business model in one act. The seller invented the constraint, the constraint is invisible, and the removal of the constraint is priced at six times the original subscription. The $200 Pro tier exists for the same reason. Nobody needs more intelligence per query. People need more queries, and the queries were rationed by the same hand that sells the ration.

The buffet where the meter still runs

The Plus tier is sold as a bundle, and the bundle deserves an audit of its own, because a subscription's list of features is not the same as a list of things that work. Sora generates video. Deep Research compiles reports. Codex writes code. Agent mode, per PCMag's February 2026 review (the review is here), does something for you rather than something with you, and the reviewer's experience of it was, in his word, disappointing, which matches mine and everyone I've compared notes with: the demos run clean, and the actual errands come back half-finished in ways that cost more time to verify than the task would have taken to do. The bundle's function is not to be used. The bundle's function is to make the twenty-dollar line item look like a bargain against the sum of its parts, so that the renewal decision is never a decision about the one thing you actually use. Meanwhile the genuinely best model, GPT-5.5 Pro, sits on the $100 tier, which means the product's own admirers are being told, gently, every month, that what they bought is not the top.

Consider the incentive, because the free tier just changed

The change that should reorganize how you think about the ladder is this: the free and Go tiers now carry advertising in the United States, per the pricing trackers that verified it against OpenAI's pages in mid-2026. Ads, on a chat product. The person who doesn't pay is not a guest, and a person who doesn't pay is inventory, the same arrangement television made with all of us and that search spent two decades perfecting. Follow the incentives down the tiers. OpenAI's costs are compute, and compute does not shrink when you subscribe. The company has raised tens of billions on the premise that it someday will. The free tier exists to grow the base, the ads exist to make the base pay, the $20 tier exists to habituate the base to billing, and the $100 and $200 tiers exist because some fraction of the base will hit a ceiling they cannot see and cannot verify and will pay to stop thinking about it. Every rung is load-bearing. My own incentives are on the table too: I pay for two of these subscriptions, I resent both, and I have not cancelled either, which is a confession about my own meter.

What the case for Plus gets right, and where it stops

The case for Plus is better than its marketing, and it deserves a real paragraph. Twenty dollars is under a dollar per workday for the most capable general-purpose tool ever sold at retail. The bundle is wide enough that one subscription replaces three or four people were already paying for. The alternative, stitching together an API key and a per-token invoice, trades a flat cost for a variable one and saves money only for light users, which the honest trackers say plainly. PCMag keeps Plus as its flagship recommendation for individuals. An analyst I know does half her first-draft work inside it and considers the price trivial against her hours. Both things are true: the meter is the product, and the product is worth what the meter charges, for the person who hits the ceiling daily. What the counterargument cannot rescue is the opacity. A fair price for a good tool does not require the price's own terms to be unreadable, and OpenAI has chosen opacity in every place it could, from the "may vary" footnote to the seat it invented a limit for and then unbilled.

The verdict, then, is conditional in a specific way. If you hit the free tier's cap on most days, pay the twenty dollars and stop reading pricing essays. If you don't, the free tier is now an ad-supported product, and your attention is the subscription you didn't choose. The variable worth watching is not the models. It's whether the caps ever become visible. A meter you can read is a tool. A meter you can't read is my old copy code, and I never once saw that code print a number I was allowed to check.