Six weeks after telling its heaviest users they would get less, OpenAI told them they would get more. Late Thursday the company said it would double the usage allowances on its ChatGPT Pro and Business Premium tiers and cancel a reduction scheduled to take effect Friday for existing subscribers to the $200-a-month Pro plan.
The change was announced on social media by Thibault Sottiaux, who runs Codex and ChatGPT. Under the new schedule, the allowance on the $100 Pro tier rises from five times the Plus limit to ten times; the Plus plan is unchanged. The reduction that had been set for Friday is gone.
Quotas are the least discussed and most consequential number in the consumer AI business. A subscription price is a headline; the limit attached to it is the actual product. When a paying customer hits a cap, the service they bought stops working in the middle of a task, and the decision they make next is whether to upgrade, wait, or leave. OpenAI’s September proposal would have lowered that ceiling for the users who rely on it most, and the reaction was loud enough that the company reversed course rather than absorb the churn.
The timing is not incidental. The reversal lands as OpenAI’s computing commitments have reached a scale that would have been difficult to describe two years ago, with multi-year agreements across chip suppliers and cloud providers that run into the hundreds of billions of dollars. It also lands amid persistent reporting that the company is preparing for a public listing. Neither fact appears in the announcement. Both bear on it. A company about to sell shares to the public wants growing usage and satisfied subscribers on the page, and generous limits produce both.
The economics of generosity depend on what is being consumed. A chat message costs a fraction of a cent. An agentic coding session, in which a model reads a repository, writes code, runs tests, reads the failures and tries again, can consume millions of tokens in an hour and keep going until it is stopped. The heaviest users of Codex are not sending messages. They are running processes, and the cost of those processes scales with success: the more useful the tool, the more of it people use.
That is why the limits exist in the first place, and why they are hard to set. Set them too low and the product is unusable for the work it was built to do. Set them too high and a small number of accounts can consume a meaningful share of capacity. The right answer changes as inference gets cheaper, and inference has been getting cheaper quickly, which is the strongest argument that OpenAI can afford this month’s increase without a matching price rise.
There is a competitive dimension as well. Anthropic, Google and a cluster of open-weight providers have all been pushing coding tools with their own usage terms, and the buyer for these products is unusually price-sensitive because the buyer is often a developer with a credit card and three browser tabs open. A cut to a limit is easy to announce and hard to explain, and the switching cost in this market is a login.
The September decision looks, in retrospect, like a test of how much friction the customer base would accept. The answer came back quickly. Subscription businesses learn this lesson repeatedly: the users most affected by a limit change are the ones generating the most usage, and those are also the ones whose word-of-mouth recruits the next cohort. Cutting them off saves money in the quarter and costs it in the year.
What the reversal does not resolve is whether the current allowances are sustainable. OpenAI has not published the compute cost behind a Pro subscription, and it does not disclose how many subscribers sit at each tier. Without those figures, a doubled quota is an assertion that the company can serve it profitably or that it can raise the money to serve it anyway. For a private company spending at this scale, those two propositions are not the same.
The likely next step is another adjustment. Limits on this kind of product are a dial, not a contract, and every provider has moved them repeatedly as costs and competition shifted. What the past two days established is the direction of the political pressure: usage limits can be raised without much explanation, but they cannot be lowered without one. That asymmetry will shape how OpenAI prices every tier it launches from here.
For subscribers, the practical effect is straightforward. The caps they were bracing for on Friday are gone, the allowance on the mid-tier Pro plan doubled, and the ceiling they hit last month now sits twice as far away. For everyone else watching the AI business, the episode is a small, clean demonstration of where the balance of power has moved. The company with the models needed the customers more than the customers needed the models, at least this week.
