OpenAI stops selling $200 ChatGPT Pro, seven days after GPT-6 Astra shipped
Existing accounts and the API are untouched, which says the constraint is long-context prefill rather than demand in general.
OpenAI stopped selling new $200-a-month ChatGPT Pro subscriptions on September 10, seven days after GPT-6 Astra shipped. Existing Pro accounts keep working.
Every other plan, and the API, stayed open.
To make sure our current users have an incredible experience and continued access to Astra, we are going to pause subscriptions to our $200 Pro plan. These put the most strain on our systems and we wanted to take the smallest step that allows us to continue giving the broadest access possible. All other plans and the api remain available.
There is no impact to existing accounts and we are working on adding more capacity as fast as we can. Thanks!
Tibo Sottiaux, who works on Codex and ChatGPT at OpenAI, posted that at 14:17 Eastern. Read the ordering rather than the apology. The company closed the top of its consumer funnel and left everything else untouched, which is the opposite of what a company does when it wants the news to be "we are wildly popular".
If this were scarcity theatre you would throttle somewhere cheap. Rate-limit Plus. Slow image generation (nobody churns over a slow picture). Quietly lower the Pro reasoning effort and let people grumble on Reddit. You would not stop taking $200 a month from people queuing to hand it over, and you certainly would not leave the API — the thing every enterprise buyer watches — running at full tilt.
So what is actually scarce?
Probably prefill, and there is outside evidence for that. Epoch AI has been measuring time to first token against context length across frontier models, and found GPT-5.6 showing a noticeable quadratic component while Claude models stayed closer to linear. On September 10 they reported that Astra follows the same shape, and that OpenAI's API price for it steps up past 272,000 input tokens.
Earlier this week, we showed that GPT-5.6 and Claude 5 exhibit very different long-context TTFT scaling, which corresponds to their different pricing structures at long context lengths.
OpenAI recently released GPT-6 Astra, which shows the same pattern of increased API pricing beyond 272k input tokens. We collected additional latency measurements, which show a similar curvature to that of GPT-5.6 models.
Now look at what Pro is on OpenAI's own pricing page. Five times the usage, "Pro reasoning powered by GPT-6 Astra", maximum Codex tasks, maximum deep research, "maximum memory and context". Pro is, almost by construction, the tier whose users push the longest contexts through the most expensive reasoning setting. If the cost curve bends upward with context, the cheapest capacity you can buy back is the marginal Pro subscriber — and you buy it by not selling the seat.
Our read: this is a compute constraint with a fixed price on the wrong side of it, and the $200 flat rate is the part that breaks. A subscription that promises "unlimited subject to abuse guardrails" against a workload whose serving cost grows faster than linearly in context is a bet that heavy users stay rare. Astra appears to have ended that. We would expect Pro to come back within roughly a month, and to come back changed — a higher headline price, a metered tier above it, or explicit context limits on the unlimited plan. If it reopens unchanged at $200 before October 10, we were reading a capacity blip as a pricing problem, and we will say so.
The generous reading deserves its hearing too. Launch weeks spike; a lab that oversells a tier and then degrades it for existing subscribers deserves more criticism than one that stops selling, and Sottiaux says explicitly that this was "the smallest step" available. Fair enough. But seven days after a launch is not a spike any more, it is a level — and OpenAI's own page was still offering a Get Pro button hours after the post went up (we checked).
Worth remembering who the pause does not affect: the people already paying. Everyone else can wait, or spend the same $200 on roughly 330 million output tokens from DeepSeek's new Flash model at its off-peak rate. Different product, obviously. Same month.
Sources
Show all 9 sources
Anthropic's threat report: 189.9m distilled exchanges, and two Chinese labs that relayed their own users to Claude
The September 10 report names seven China-based labs, a Russian espionage actor whose agents rebuilt malware until it went undetected, and customer data that reached Anthropic by accident.
Continue ↓