What Will Happen to Our Access to ChatGPT and Claude After Their IPOs. The Main AI News
Will an ordinary user keep access to the strong models once the state gets them first and OpenAI and Anthropic, after their IPOs, have to earn on tokens instead of giving them away with a subscription?
The White House order of June 2 made model review voluntary — 30 days instead of a mandatory 90 — but, as Ilnar Shafigullin shows, the state still gets the model a month before release along with its own benchmarks, and Anthropic is widening access to vulnerability hunting from 50 to 200 companies. Alexander Volchek does the arithmetic: analyzing one text through the API cost 50 dollars, his work in ChatGPT is worth a hundred thousand a month, and Codex has only 5 million active users, whom paid tokens would cut down to "you and me". Ilnar explains how Dynamic Workflow works in Claude 4.8 and why he increasingly works in Codex rather than browser ChatGPT, and the host answers why he himself has not switched. The episode's key shift is Anthropic's bet not on coding but on software for the whole world, on the Microsoft model, and the advice to look not at the IPO but at those these companies will be paying.
What to watch for
Key takeaways
Trump pushed a bill under which every American model would have to give the government access 90 days before release; it was shelved, but on June 2 the White House issued an executive order with a voluntary review mechanism 30 days before release. Sam Altman spoke out against mandatory government permission, which puzzles the host given Anthropic's conflict with the state after the capture of Venezuela's president and the contract OpenAI signed the very next weekend. Alexander Volchek's conclusion: the state will get effectively unlimited use of the models, while users will see fewer and fewer models in real access.
Ilnar Shafigullin explains: the model decides which subtasks to split a task into, changes the plan as it goes, writes a small prompt with context for each subtask, and agents — small ones or other instances of the same model — work from it. A person rises one level — from chatbot operator to the one who sets the task for a model-orchestrator — and "tokens flow like a river". Alexander Volchek has not seen anything exclusive in his own systems yet: asked about special features, the system itself says there are none.
Working through the API, the host saw that every system is set up not to report how tokens are spent: analyzing one text cost 50 dollars, and he values his work in ChatGPT at a hundred thousand dollars a month. He recalls Jensen Huang's line: a programmer on a 500-thousand salary should spend no less than 500 thousand on tokens. Cheap models turn into errors — Zoom's summaries do not match the conversation, whereas with ChatGPT 5 Pro an hour of transcription would cost at least 5 dollars.
The host has an enormous number of chats and projects, and ChatGPT has started merging information between projects that must not be merged; the command "act strictly this way" does not always work. Claude 4.8 Max, asked to do an audit, writes straight into the repository and then answers "sorry, I made a mistake". Ilnar suggests separate accounts for separate tasks; the host is unsure it is legal and is ready to overpay for a single system with memory.
A friend of the host claimed Russia leads in robots thanks to MIPT's walking robot, but a video from the St. Petersburg economic forum of MIPT robots playing football, next to the project of Boston Dynamics, owned by Hyundai, shows the gap. That, Alexander Volchek says, is how free versions and open source models differ from what governments are discussing. The Chinese government, he adds, has been checking models in advance for fifteen years and will pass no order about it.
Ilnar opens the browser version of ChatGPT less and less and works more and more in Codex. Alexander Volchek has not switched to Codex, although he builds large infrastructure projects there: moving all his requests would mean paying for tokens without limits, and data storage is awkward — fresh chats have vanished on him about fifteen times, old ones get erased, and there is no multi-repository GitHub work. ChatGPT has no limits, and he finds 5.5 Pro and 5.5 Thinking far stronger than 5.5 High in Codex.
The host has never opened the code of his AI-native systems, never uploaded or committed anything, so he does not call it vibe coding: the code is written by AI for AI. Ilnar agrees that Karpathy's term has become a meme, and in his circle it is used for analysts and product managers who sent a request and got a result. The one-person unicorn, he says, is closer: an MVP can certainly be built now without technical specialists.
Several months ago employees of state companies and institutes in China could travel abroad only with written permission, and a few weeks ago the rule was extended to commercial companies where an important specialist works. Ilnar compares it to exit lists in the Soviet Union: scientists have become a strategic question for whole countries. Alexander Volchek recalls the Manus story: the company was sold to Meta, the assets moved to Singapore, and the founders were not let out.
The original wording obliged companies to submit their models for testing; Silicon Valley sold its own version, and the order states outright that this is not licensing. But API access is strongly recommended, the state creates its own benchmarks and gets the model 30 days ahead. Ilnar links this to Claude Mythos and Project Glasswing: Anthropic is widening access to vulnerability hunting from 50 to 200 companies, and the state apparently joins that list ahead of public release.
Ilnar believes OpenAI now gives the models away effectively free for the sake of market share, and after the IPO, when the shares are sold and it is time to earn, subsidizing users may give way to token charges; the same awaits Anthropic. Alexander Volchek adds the goal of a listing — a valuation of ten trillion rather than one — and recalls that today nobody is allowed into these companies even at two trillion. The winner will be whoever has an audience of a billion plus, so losing users for the sake of profit is a bad sign.
If Claude Code builds a CRM for 100 million people or a messenger for a billion, people will use that software, not Anthropic — just as billions use Microsoft's Windows, Word and Excel. The host himself has stopped using Excel: ChatGPT creates and analyzes the tables. Codex so far has only 5 million monthly active users, and paid tokens would cut them down to "you and me", yet the host would be glad of a ChatGPT Pro at a thousand dollars — for the wider access OpenAI hides behind 200.
The equipment is installed at home, connected to the owner's electricity, and all the bills are paid, while the station works as a server — an answer to the protests against data centers opposite residential homes, which many states ban by decree. Ilnar compares it with the Liberman brothers' idea of distributed computing. The host considers the project risky: hardware becomes obsolete fast, big facilities are easy to repurpose and a network like this is not, and he would not invest.
What this episode is about
The ToTheMoon Sunday podcast with Alexander Volchek, Ilnar Shafigullin and Tatiana Tsvetkova. The occasion is a week in which the White House order on model reviews, the Claude 4.8 update, new token bills and the question of what happens to access after the OpenAI and Anthropic IPOs all landed at once.
The first block is the state and the models. Trump pushed a bill giving the government access to models 90 days before release; it was shelved, but on June 2 the White House issued an executive order with a voluntary 30-day review mechanism. Alexander Volchek recalls Anthropic's conflict with the state after the capture of Venezuela's president and the contract OpenAI signed the very next weekend, and concludes: the state will get effectively unlimited use of the models, while users will get fewer and fewer models in real access. Ilnar later adds the details of the order: Silicon Valley sold its own version, it does not count as licensing, but API access is strongly recommended, the state will have its own benchmarks, and Anthropic is widening access to vulnerability hunting through Claude Mythos and Glasswing from 50 to 200 companies.
The second block is Claude 4.8 and the price of work. Ilnar explains Dynamic Workflow: the model splits a task into subtasks, writes the prompts and runs the agents itself, and a person rises from chatbot operator to someone who sets the task for a model-orchestrator; the host has not felt anything exclusive yet. Then come the tokens: analyzing one text through the API cost 50 dollars, the host values his work in ChatGPT at a hundred thousand dollars a month, recalls Jensen Huang's line about a programmer on a 500-thousand salary, and explains why Zoom's summaries do not match the conversation. Separately: ChatGPT, which has started mixing independent projects, and Claude 4.8 Max, which writes into the repository unasked and apologizes.
The third block is tools. A video of MIPT robots playing football next to the Boston Dynamics project becomes a metaphor for the gap between free versions and the top models. Ilnar admits he increasingly works in Codex rather than browser ChatGPT; the host explains why he has not switched: without subscription limits you pay for tokens, chats disappear, there is no multi-repository GitHub work, and he finds 5.5 Pro and 5.5 Thinking stronger than 5.5 High. The argument about vibe coding ends in agreement that Karpathy's term has become a meme, and the talk of a one-person unicorn ends in the relay practice of passing specs between Claude and Codex, which a study calls a mistake.
The fourth block is borders and hiring. China has extended its exit ban for key specialists from state companies to commercial ones; Ilnar compares it to exit lists in the Soviet Union, and the host recalls Manus, sold to Meta with its assets in Singapore and its founders unable to leave. The host considers ChatGPT's new job search a backward feature and instead describes a Codex service that crawls the portals itself, matches vacancies against a résumé, syncs with a calendar and messages you on WhatsApp or Telegram.
The fifth block is the IPO. Ilnar expects we will start paying for tokens after the listing, when market share gives way to profit. The host explains the goal of an IPO — a valuation of ten trillion instead of one — recalls that today nobody is allowed into these companies even at two trillion, and formulates the main shift: Anthropic's bet is not on coding but on software for the whole world, as Microsoft's is on Windows, Word and Excel. Codex has only 5 million active users, paid tokens would cut them down to "you and me", and yet the host would be glad of a ChatGPT Pro at a thousand dollars for the sake of wider access. To investors he recommends looking not at the IPO but at the contractors: the xAI contract with Anthropic is a billion dollars a month.
The episode closes on money and hardware: Google selling shares for 80 billion for AI, the cash that sat idle at Apple fifteen years ago now going into servers and electricity, and the NVIDIA startup that places 200 thousand dollars of equipment in homes for 150 a month — an answer to the protests against data centers that the host considers risky.
The episode is useful because it ties scattered news into a single line of access: who gets the model first, what it really costs and what changes once the companies have to earn. The hosts do not argue with the inevitability of paid tokens — they show where the user is already paying without knowing it, and where to look when the IPO happens.
The episode's main point is that access to the strong models has already stopped being a given: the state gets them first and without limits, the cost of work hides in tokens the systems do not show, and an IPO will move the companies from capturing share to earning profit. The hosts do not scare, they count — 50 dollars per text, 5 million Codex users, 300 billion in raised money — and suggest looking not at the listing but at those who will receive that money. Alexander Volchek's forecast that Anthropic is betting on software for the whole world rather than on programmers is the thesis worth testing against time.
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 250 segments: 0 identified, 0 mixed, 250 probable, and 0 unresolved.
Read transcript on a separate page
Loading…