The most expensive variant of ChatGPT Pro’s subscriptions, which is normally available for 200 dollars, is temporarily unavailable for new sign-ups. The company’s product lead Thibault Sottiaux confirmed the pause after suggesting its possibility earlier in the week. The rampant demand for GPT-6 Astra and the heavy load on OpenAI’s infrastructure for Pro 20x in particular are the immediate reasons. However, this looks like the shape of things to come for ‘subsidized’ AI.
It should be noted that current Pro subscriptions are unaffected. The option to upgrade to Pro at the 200 dollar per month tier has simply been taken away, seemingly temporarily.
ChatGPT Plus (20 dollars per month) and Pro (either $100/m for 5x usage or $200/m for 20 times Plus’ usage) provide far more access to OpenAI models than the same monetary amounts would offer through API access. For this reason, many call these ‘subsidized’ tiers compared to the ‘real’ costs through API-billed tokens. Although the chosen LLM and its use case differ, it is indeed trivial to rack up thousands of dollars in API costs for what would otherwise cost you 100 or 200 dollars for the Pro subscription.
(Not) the end
Ever since ChatGPT’s launch, users have been able to cost OpenAI enormous sums of money. ChatGPT Plus, introduced in February 2023 mere months after the chatbot’s launch, simply used the 20 dollar subscription tier amount for mass-market appeal. When Pro was launched in late 2024, it was in response to power users requiring far more usage than Plus allowed for; some resorted simply to buying multiple subscriptions to get more access. Pro at the 20x tier in particular has been highly lucrative to the end user, which selects for those putting the most strain on OpenAI’s infrastructure.
For now, subscriptions for Pro at the 20x/200 dollar tier are merely on pause ‘for a bit’, according to Sottiaux. Nevertheless, OpenAI’s main competitor Anthropic has already tried hard to make its cheapest subscription tier less appealing, while also putting a rather strict cap on usage even for Max 5x and Max 20x users. Fable 5 and 5.1 are only available on the latter two subscription tiers, and even there only for 50 percent of the total usage allowed within the plan.
A race to the top
Equivalent intelligence, however you wish to benchmark it, has gotten cheaper by drastic amounts. Any LLM at the pricey frontier in mid-2025 is now bested by 2026 open-weight models and mid-tier offerings by the leading AI labs. However, most users will inevitably flock to the latest and greatest LLMs. This is shown by the fact it wasn’t the efficient and broadly available GPT-5.6 Luna or Terra, but the seemingly gargantuan GPT-6 Astra that made OpenAI buckle.
There is little doubt the capabilities found in Claude Fable and GPT-6 Astra will be bested once more by cheaper offerings down the line, be it in 6, 12, or 18 months. The point is that even with that being the case, these two main AI players racing ahead to new frontiers means that their infrastructure will remain at a premium, demand and scale soaring beyond capacity virtually at all times nowadays.
This means that the subscription tiers are a pain in the backside for both. It seemed, with OpenAI raining down treasure troves of weekly resets to boost value for money for subscribers, that the subsidization would continue apace. Now, there’s more than a grain of doubt creeping in.
Expect silos
We should expect further silos of more capable models. Claude Pro already isn’t the product it used to be now that the frontier LLM (currently Fable 5.1) is elusive, and OpenAI was even unsure about rolling out GPT-6 Astra to Plus users quickly, instead committing only to try to provide access. There’s no need to assume malicious intent here: these labs have allegedly always operated at a loss, and with them being private companies, no financials are yet available in detail to suggest otherwise.
In order to protect not only margins through API calls but also capacity for the most important and lucrative users (those being enterprises and the public sector over individual developers and other private individuals), both the cheapest and most expensive tiers are expected to collapse in on themselves. Expect the lowest tier (e.g. Claude Pro and ChatGPT Plus) to remain hobbled in some way (side-note: even though Astra is technically available at the Plus tier, usage is extremely constrained). Moreover, expect the 20x tiers, offering an order of magnitude more usage than the bottom subscription for ‘only’ 10x the price, to be toned down or trade their sizable usage limits for exclusive access to better models. The economics never worked, and the immense capacity constraints are expected to only worsen as model scale, adoption and the need to race to pre-IPO profitability all push these tiers to one side.
Also read: GPT-6 Astra, OpenAI’s answer to Fable, begins rolling out