OpenAI has stopped selling new ChatGPT Pro subscriptions, its $200-a-month top tier, because demand for the Astra model is outrunning the compute behind it. Thibault Sottiaux, the OpenAI product lead who oversees Codex and ChatGPT, announced the pause on X, saying Pro puts the heaviest load on the company's systems and that registration had to be switched off temporarily. The API, Go and Plus are all still selling. A little over two weeks after shipping the model it called the beginning of the AGI era, OpenAI is refusing money from the customers who pay it the most.
Sottiaux described the pause as the minimum restriction that preserves the widest possible access everywhere else. That is an accurate description of rationing. The implied arithmetic is blunt: Pro subscribers consume enough capacity that closing their queue is the cheapest single lever for keeping everyone on the lower tiers served.
The warning came first. On Wednesday OpenAI said it might have to suspend new Pro signups, with Sottiaux writing that demand for Astra had been unprecedented, that the company was using every available means to absorb the load, and that it had never encountered growth like this. Keeping service stable for existing users was the stated priority; if demand held, new Pro registrations could close for a short period. Demand held.
What the company has not disclosed is how long the pause will last or how many people were subscribing per day. Without either number, "unprecedented" is a claim that cannot be checked against anything — including OpenAI's own previous launches. Capacity announcements are usually where a company reaches for figures, not away from them, and the absence here reads less like modesty than like a number nobody wants attached to a quarter.
There is one dated marker of how recent the squeeze is. Only last month OpenAI raised usage limits on Codex, which is not a decision a company makes while it is running out of room. Whatever broke, broke after that.
Astra was introduced on 3 September and rolled out gradually to the Pro, Plus, Enterprise and Business tiers as well as to OpenAI's other AI products. The company promised material progress in three areas: reasoning, coding and computer use. All three are contested ground, and all three are among the most compute-hungry things a model can be asked to do. OpenAI also called Astra a jump between model generations and the start of the AGI era, language that did its job.
That is the part worth sitting with. The pitch and the outage are the same event. A model marketed on long-horizon reasoning, agentic coding and driving a computer invites exactly the workloads that consume tokens by the hour rather than the message, and the tier that attracts those users is the one now closed. OpenAI did not oversell a product and get caught; it sold the product accurately and got caught anyway. The gap is not between the marketing and the model, it is between the model and the data centres.
It also sets a quiet precedent. Pro exists to be the plan without limits, the one you buy so you never have to think about quota. A company that suspends sales of its unlimited tier has conceded that the tier is priced against a supply curve like everything else — and $200 a month is now demonstrably below what the heaviest users cost to serve.
Astra was announced as a generational leap. Two weeks on, the binding constraint on the era it was supposed to open is not the model's reasoning. It is how many people OpenAI can afford to let use it.