Here lies a figure that should make any accountant at OpenAI feel uneasy. A ChatGPT Pro subscription priced at $200 a month, when pushed to its limits, can deplete work valued at approximately $14,000 monthly based on the rates OpenAI charges developers purchasing the same output directly. This stark contrast between what a frequent user pays and the actual value of their usage is significant. This data originates from a June 2026 examination conducted by SemiAnalysis, an independent research organization that opted to measure the economics of these plans rather than speculate.
## How $200 converts to $14,000
The approach taken was straightforward. Rather than approximating the figures externally, the organization acquired the plans and utilized them extensively. SemiAnalysis stated, “Recently, we acquired every Anthropic/OpenAI subscription plan and conducted prolonged coding tasks until we hit the weekly cap.” Each unit of usage was then priced according to what a developer would pay to obtain the same output directly through OpenAI’s pay-as-you-go model.
What renders the outcome astonishing is how far it deviates from common assumptions. SemiAnalysis remarked, “It’s generally thought that a $200/month plan peaks at around $2,000/month worth of tokens (considering API pricing).” Developers had anticipated a limit near $2,000. The actual figure was several times greater: approximately $14,000 for ChatGPT Pro’s highest-tier 20x plan, and about $8,000 for Anthropic’s Claude Max 20x, which also has a monthly fee of $200.
The disparity enlarges as you progress through the tiers. A fully utilized $20 ChatGPT Plus plan equates to roughly $700 of usage. Paying ten times more for Pro leads to a potential subsidy that escalates far beyond tenfold. The premium plans exhibit the biggest discrepancies.
## The catch is at the ceiling
These figures stem from stress-testing, and it’s important to clarify what they represent and what they do not. This measurement comes from one firm, executed by purposely maxing out each account. It is not a peer-reviewed study, nor does it reflect typical tool usage patterns. The per-subscriber loss estimates also rely on their own assumptions, including an estimated profit margin. Interpret them as modelling from SemiAnalysis, not as definitive truths.
The $14,000 figure surfaces only if an account is excessively utilized week after week, performing prolonged, continuous coding tasks that consume usage incessantly. Very few individuals do that. The more compelling inquiry is where the break-even point stands, and the answer is disturbingly low. SemiAnalysis identified the juncture at which OpenAI begins to incur losses on its top-tier plan at merely 5.7% of the maximum permitted usage. For ChatGPT Plus and the Pro 5x plan, the cost-deficiency line hovered around 11.4%. Anthropic’s plans performed somewhat better: the firm estimated break-even on Claude Pro and Max 5x at approximately 20%, with no profit on its top tier at about 10%.
These thresholds reshape the entire scenario. A provider does not need a user to reach $14,000 to begin losing money on them. It requires only a small portion of its most active users to exceed single-digit usage, and the plan is already operating at a loss for that subset.
## Why price it this way at all
If the ceiling poses such risks, what is the rationale behind offering a flat rate? Most subscribers never come close to that limit, and lighter users effectively subsidize the heavier ones. A consistent $200 fee is straightforward to sell and easy to budget, with the vast majority of those who pay it utilizing only a fraction of what they could. The plan hinges on the premise that the average balances out.
What complicates that premise is a shift in how the tools are utilized. A single chat prompt is inexpensive. However, an “agentic” system that plans, employs tools, retries, and independently works through an entire task can necessitate up to 1,000 times more usage than a typical prompt. As more individuals apply these plans to such tasks, the comfortable gap between the average subscriber and the break-even line diminishes. This pressure contributes to the broader trend observed in 2026, moving away from unlimited pricing models toward charges based on actual usage.
Moreover, there’s an alternative economical route that heavy users and enterprises are currently pursuing. As per a Wall Street Journal article, delegating routine tasks to less expensive or free open-source models, while reserving the high-end models solely for genuinely complex problems, can reduce expenses by as much as 95%. Vishal Misra, a computer scientist and vice dean at Columbia, articulated it clearly to the WSJ: “You don’t need a model that understands quantum gravity” to reformat a spreadsheet. His wider perspective is that as capable open models proliferate, the price premium people