Skip to content
Token limits · Artificial intelligence · OpenAIEpisode 168 · 20 September 2026 · 01:04:29

OpenAI Closed Access to the $200 Pro Plan. Is Strong AI Becoming Scarce?

Central question

Is access to the strongest models becoming a scarce resource of its own — and how do you build daily work with AI when plans get closed, limits and modes change, and the companies argue over who should check their models?

What you take away

OpenAI closed registration of new Pro users at 200 dollars after the rush to GPT-6 Astra; existing accounts are unaffected, and Ilnar advises leaving the subscription alone and keeping money on the card; both hosts expect the plan to return or a new one timed to hardware or a data centre. Limits have become a factor in choosing a model: on OpenAI's Pro they are generous and reset often, on Claude they run out faster, and Anthropic has just raised the limits of its paid tiers — a possible sign of user outflow. ChatGPT no longer switches to deep reasoning on the phrase 'think long', so Alexander advises against the fast-answers mode and for working in Extra High or Pro; OpenAI has released an official instruction on writing requests for Astra, and Agents.md files worsen results with the new models. An old context window proved more valuable than a new one: a task that GPT-6 Astra and Fable 5.1 did not solve was solved in a chat older than a month, Fable itself refused to move, and Codex now confirms only complex actions. Amodei called for slowing development and admitting outside checkers, Musk backed checks of models by other companies, Trump declared there is no threat, a federal court found the ban on Anthropic in the defence complex unlawful, and Anthropic's report showed that hacking operations with Claude have become cheaper and are uncovered after the fact. Apple released Siri AI, but the host has no scenarios for it: ChatGPT already holds his entire context.

Main threads

What to watch for

1If you have a $200 ChatGPT Pro subscription and need it, change nothing and make sure the payment goes through: getting back onto the plan after the pause may not be possible.
2Do not use the fast-answers mode: pick Extra High, Pro or the very latest mode — a simple question is still answered instantly, and long reasoning means it is needed.
3Do not abandon an old context window for a new chat in the same task: context compression works better than manual transfer, and a new chat has to be configured again.
4Review long Agents.md files and old instructions for the new models — with GPT-6 Astra they can worsen results and increase token consumption.
5Choose one main model for daily work and work with it directly rather than through third-party services, where Astra Pro and Extra High are unavailable via the API.
Signals to track afterwards
Whether OpenAI brings back the $200 Pro subscription or replaces it with a new plan — and whether paid limit resets appear.
What Sam Altman says at OpenAI Dev Day in two weeks, and Alexander's report from there.
The response of Anthropic and the US government to the court ruling on the unlawful ban in the defence complex, and the course of Anthropic's IPO.
Whether real third-party checks of models, proposed by Amodei and backed by Musk and Altman, materialise.
The first Siri AI use cases from subscribers, which the host asks for.
Most useful for
Subscribers to ChatGPT Pro, Plus, Claude and Codex: what is happening with plans, limits and reasoning modes.Developers working in Codex and Claude Code: when not to open a new chat and what to do with old Agents.md files.Anyone following the politics around AI: the statements by Amodei, Trump and Musk, Anthropic's court case with the US government, and the IPO.Security specialists: what Anthropic's report on Claude in cyberattacks showed.Apple users waiting to see what the new Siri AI brings.

Key takeaways

01:29OpenAI Closed $200 Pro Registration After the Rush to Astra

Against a billion users and 70 million requests per second, the company stopped selling Pro to new users; existing accounts are unaffected, and Ilnar advises not changing the subscription and keeping money on the card.

08:07Limits Have Become a Factor in Choosing a Model

On OpenAI's Pro the limits are generous and reset often, on Claude they run out faster; Anthropic has just raised the limits of its paid tiers — a possible sign of user outflow after the GPT-6 hype.

20:16ChatGPT No Longer Switches On Deep Reasoning at the Phrase "Think Long"

The fast-answers mode is a saving for OpenAI; Alexander advises working in Extra High or Pro, where a simple question is answered instantly and long reasoning means it is needed.

24:10An Old Context Window Is Worth More Than a New One

A task that GPT-6 Astra and Fable 5.1 did not solve was solved in a chat older than a month, and Fable refused to move; Ilnar compares manual context transfer with prompt engineering, which got built into the systems.

34:58Amodei Asks to Slow Down, Trump Answers: There Is No Threat

The Anthropic chief's statement on independent checks dropped the AI market, and Trump, by phone at the AI Summit, declared the threat invented and AI worth developing; Alexander reads it as moves ahead of Anthropic's IPO.

39:04Musk Backed Checks of Models by Other Companies

Musk suggests his models be checked by Google, Meta, OpenAI and even Chinese companies; according to Reuters, Amodei's call was backed by Musk and Altman, and a court found the ban on Anthropic in the US defence complex unlawful.

46:37Anthropic's Report: Hacking Operations With Claude Have Become Cheaper

Anthropic showed how hacker groups use models a class below Fable and Mythos; per Ilnar, a few enthusiasts of average qualification can now cause serious trouble, and it is uncovered after the fact.

59:32Siri AI Is Out, but There Are No Scenarios for It

Siri AI searches messages, mail and photos, but Alexander keeps Siri switched off: ChatGPT already holds his entire context, and Apple Photos search already finds "Polina with a fish"; the iPhone 18 is an automatic purchase with no AI expectations.

What this episode is about

The Sunday ToTheMoon episode with Alexander Volchek and Ilnar Shafigullin opens with OpenAI's numbers — more than a billion users and 70 million requests per second — and with the closing of Pro registration at 200 dollars. ChatGPT explains it by server load and demand for Astra, Alexander suspects a pricing mistake and repeats his rule: buy the paid version, preferably Pro — it is not a Netflix subscription but your professional development. Ilnar relays an OpenAI employee's tweet that existing accounts are unaffected and explains why limits have become a factor in choosing a model: OpenAI's are generous, Claude's run out faster, and Anthropic has just raised the limits of its paid tiers — possibly having noticed an outflow of users after the hype around GPT-6.

Next — how working with the models itself has changed. ChatGPT no longer chooses deep reasoning for the user, so the phrase 'think long' has stopped switching the model; Alexander advises against the fast-answers mode and for Extra High or Pro, where simple questions are still answered instantly. OpenAI has released an official instruction on writing requests for GPT-6 Astra, and a study showed that Agents.md files worsen results with the new models and increase token consumption. Alexander tells how a task that GPT-6 Astra and Fable 5.1 did not solve was solved in a context window older than a month, and Fable refused to move to a new chat; Ilnar compares this with prompt engineering and orchestrators that got built into the systems, and advises using the settings out of the box. Both stay on ChatGPT as their main model, Alexander works more and more in Claude Code, and considers bots through Excel and Word a topic for those who do not work with a computer.

The second half is the politics around safety. Dario Amodei called for slowing model development and allowing outside checks, the AI market fell, and Trump, by phone at the AI Summit, told Jensen Huang and the hall that there is no threat and AI must be developed. A federal court found the ban on Anthropic in the US defence complex unlawful, Elon Musk backed checks of models by other companies, including Chinese ones, and an Anthropic IPO is being discussed at a valuation of 2–4 trillion. Ilnar goes through Amodei's text — the motif of democratic countries and the concrete points on outside access — and Anthropic's report on Claude in cyberattacks, whose main conclusion is that hacking operations have become cheaper and are uncovered after the fact. Alexander speaks of the opened 'Pandora's box', of the memes about a 10 percent probability, and of hype in which sound decisions are hard to make. The finale is Siri AI on the generation-27 systems, for which the host has no scenarios because ChatGPT already holds his entire context, and the wait for the iPhone 18.

The value of the episode is that the closing of one plan is taken apart as a symptom: access to the strongest models is becoming a resource distributed through limits, modes and prices, and the hosts show it on their own token counters and on the changed behaviour of Codex, Fable and ChatGPT. The political part — Amodei, Trump, Musk and the court — is presented without taking sides: the statements are read as market moves ahead of an IPO, not as forecasts. The practical conclusion: stick to one main model, do not economise on the reasoning mode, and do not transfer context by hand — the system does it better.

Episode transcript

The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 127 segments: 127 identified, 0 mixed, 0 probable, and 0 unresolved.

Read transcript on a separate page

Loading…