Here is a number that ought to make any accountant at OpenAI wince. A $200-a-month ChatGPT Pro subscription, pushed as hard as its limits allow, can burn through work worth roughly $14,000 a month at the prices OpenAI charges developers who buy the same output directly. That is a huge gap between what a heavy user pays and what their usage is actually worth. The figure comes from a June 2026 test by SemiAnalysis, an independent research firm that decided to stop guessing at the economics of these plans and just measure them.

How $200 turns into $14,000

The method was blunt. Instead of estimating the numbers from the outside, the firm bought the plans and ran them into the ground. SemiAnalysis wrote that “Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit.” Every unit of usage they consumed was then priced at what a developer would pay to buy the same output directly through OpenAI’s pay-as-you-go pricing.

What makes the result surprising is how far it sits from what people assumed. SemiAnalysis noted that “It’s widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing).” Developers had figured the ceiling was around $2,000. The real number was several times higher: roughly $14,000 for ChatGPT Pro’s top-tier 20x plan, and around $8,000 for Anthropic’s Claude Max 20x, which also costs $200 a month.

The gap grows as you move up the tiers. A fully-used $20 ChatGPT Plus plan is only worth about $700 of usage. Pay ten times more for Pro and the potential subsidy grows far faster than tenfold. The premium plans are where the numbers get most lopsided.

The catch is at the ceiling

These are stress-test numbers, and it is worth being clear about what they are and aren’t. This is one firm’s measurement, done by deliberately maxing out each account. It is not a peer-reviewed study, and it is not how anyone normally uses these tools. The per-subscriber loss estimates also rest on assumptions of their own, including an assumed profit margin. Read them as SemiAnalysis’s modelling, not settled fact.

The $14,000 figure only shows up if someone drives an account into the ground week after week, running the kind of long, continuous coding tasks that eat through usage nonstop. Almost nobody does that. The more interesting question is where the break-even line sits, and the answer is unnervingly low. SemiAnalysis put the point where OpenAI starts losing money on its top-tier plan at just 5.7% of the maximum allowed usage. For ChatGPT Plus and the Pro 5x plan, the money-losing line was around 11.4%. Anthropic’s plans held up a bit better: the firm estimated break-even on Claude Pro and Max 5x at around 20%, and no profit on its top tier at roughly 10%.

Those thresholds flip the whole picture. A provider does not need a user to hit $14,000 to lose money on them. It only needs a small slice of its heaviest users to push past single-digit usage, and the plan is already underwater on that group.

Why price it this way at all

If the ceiling is this dangerous, why offer a flat rate? Because most subscribers never come close to it, and the light users pay for the heavy ones. A predictable $200 charge is easy to sell and easy to budget for, and the vast majority of people who pay it use a sliver of what they could. The plan is a bet that the average holds.

What is straining that bet is a change in how the tools get used. A single chat prompt is cheap. But an “agentic” system, one that plans, uses tools, retries, and works through a whole task on its own, can require up to 1,000 times more usage than a standard prompt. As more people point these plans at that kind of work, the comfortable gap between the average subscriber and the break-even line shrinks. That pressure is part of why 2026 has seen a broad shift away from all-you-can-eat pricing toward billing by how much you actually use.

There is also a cheaper path that heavy users and companies are already taking. According to a Wall Street Journal report, sending routine tasks to cheaper or free open-source models, and saving the top-end models only for genuinely hard problems, can cut costs by up to 95%. Vishal Misra, a computer scientist and vice dean at Columbia, put it plainly to the WSJ: “You don’t need a model that knows quantum gravity” to reformat a spreadsheet. His broader view is that as capable open models spread, the premium people pay for top-tier access will likely shrink, though that is a prediction, not a settled outcome. SemiAnalysis itself suggested that as the underlying costs fall, top-tier access could eventually be sold at a profit for something closer to $20 a month.

Who these plans are really built for

The uncomfortable truth in this experiment is that the sticker price of an AI subscription tells you almost nothing about its value, and the plan is built around that fact. The $200 charge is set for the typical subscriber who barely touches the limits, not for the engineer running agents around the clock. The heavy user is getting a real bargain, paid for by everyone who buys the same $200 plan and uses a fraction of it.

The number that matters most here is not the $14,000. It is the gap between what you pay and what you actually get out of it. For most people that gap runs the other way: they are paying for a ceiling they will never reach, funding the handful of users who live near it. If your own usage is modest, you are not the person the pricing fears. You are the reason it can exist at all.