ChatGPT Will Never Be the Same? What OpenAI Is Preparing
What will happen to the familiar ChatGPT once OpenAI merges it with Work and Codex and removes unlimited use — and why does AI call not for automating old processes but for building new ones?
OpenAI, judging by an email Alexander Volchek received right on the road, will merge ChatGPT with Work within one to two weeks and remove unlimited work in the chat — the kind on which, he says, even the five-hundred-dollar Pro plan let you work with Astra without limits and upload tens of thousands of pages, a million calls and hundreds of thousands of tickets to take a company apart before coming in as a shareholder. Alexander puts the split into ChatGPT, Codex, Work and Dots down to the narrow view of the labs' project managers, whose leaders themselves do little work in their own systems: Greg Brockman praises a blogger's app for interview preparation, although Codex, from a single spoken prompt — download seven years of interviews, transcribe them with a local Whisper, comb through a thousand sources and break the work into five to ten stages on GPT-6 Astra Max — will assemble such a dossier on its own, and even that Alexander calls the old world of AI. The main mistake goes back to the era of ERP and CRM: companies automated their existing processes by asking employees, instead of changing them — while AI makes it possible to build systems that don't have a name yet and makes hundreds of thousands of startups solving a local task unnecessary.
What to watch for
Key takeaways
OpenAI plans to merge ChatGPT and Work within one to two weeks. Its system now consists of ChatGPT, Codex, Work and Dots, with a line of separate systems alongside: one just launched for biologists, one for education and, as Alexander believes, one for the military; meanwhile a strong project for researchers is being wound down. Many people, he notes, don't even know what Work is.
Codex was conceived as a system that works together with your computer, and that is exactly what confuses people: they think it's for geeks and programmers. In reality, Alexander says, Codex is a new format and an upgrade of ChatGPT: recurring tasks work better, and working with files, reports, presentations and data processing is more convenient.
In ChatGPT you choose a default model and its working level, and the medium mode may put on GPT-5.6 Sol and the maximum one Astra. Codex has the modes Medium, High, Max and Ultra Code but no notion of Pro. From the very start Codex worked differently within a session: as if it searched the internet worse and didn't analyze the common base the same way, as though the model were deployed only locally.
OpenAI conceived Work as an environment for working with documents, files, presentations and email — without programming, which stays with Codex. Alexander considers it a story copied from Anthropic's Cowork and recalls that Cowork has disappeared at Anthropic too.
The OpenAI and Anthropic teams, in Alexander's view, didn't understand what their systems would achieve, and their leaders rarely use Codex the way advanced people in business do. "Where would Sam Altman find so much time to sit and figure out how to build various processes?" Without doing it himself, he isn't living through the new era of use cases.
In an interview Alexander's friend was watching, a blogger showed Greg Brockman, OpenAI's number two, an app that gathers information about a guest before a meeting and prepares questions, and Brockman called it a great system. "What great system?" Alexander responds: today the same thing is done in Codex with a single request.
Alexander's prompt: download seven years of Brockman's interviews from YouTube, transcribe them with a local Whisper, comb through no fewer than a thousand sources and social media, compile the trends, failures, right and wrong predictions, conflicts and his position on the twenty main questions about AI, then go through the author's own interviews and suggest questions in his style. Such a prompt, he says, can be copied and pasted into Codex — and the same can be done for any person.
So that Codex doesn't make mistakes, the work should be broken into no more than ten and no fewer than five stages, start with a general audit and use GPT-6 Astra in Max mode. That doesn't mean the model will do everything in five or ten runs: the execution will have to be adjusted along the way. And even such a task, Alexander adds, is already the old world of artificial intelligence.
Since 2003, when Alexander got hooked on data normalization, he worked on enterprise automation — ERP, CRM, various systems. The fundamental mistake he always saw: processes were automated without being changed. Contractors came to management and then went to the employees, and they wanted to make work easier for themselves, not to help the company earn more.
Great automated systems, Alexander says, are created from scratch — like Airbnb or Uber: someone comes with a new idea and sets a new business process for the market, and old companies gradually catch up with it.
AI makes it possible to bring changes into a company that differ colossally from standard ERP-, CRM- and MRP-style automation, and to build systems whose names haven't been invented yet. What to call a system that assembles a dossier for an interview is unclear, and in Alexander's view even those who consider themselves advanced won't carry such a task through to the end.
OpenAI and Anthropic didn't foresee such cases from the start. The core of the models is built by the most brilliant teams, but the projects around them — mailings, websites, learning systems — look as if they were built in the 2010s. Alexander explains the split into ChatGPT, Codex, Work, Claude Code, Cowork, Claude Design, Claude Duet and Dots by the narrowed perception of project managers who recreate processes already described on the market.
It's not about expensive models but about thinking: would you have been able to launch such a case in ChatGPT two years ago and foresee it? Tens — actually hundreds — of thousands of startups developing systems for AI have already died in the world, and more will die: they solve a local task that AI will solve anyway. Only a tiny minority, Alexander says, may get lucky — with the right arrangements with investment funds.
OpenAI plans to remove the notion of unlimited work. Unlike Anthropic, ChatGPT kept good models essentially without limits, and the five-hundred-dollar Pro plan, Alexander says, was unlimited even on Astra. His cases rested on that: uploading tens of thousands of pages, a million calls, two hundred thousand tickets, marketing and financial data to analyze a company he joins as a shareholder or on the board.
What this episode is about
A solo ToTheMoon episode that Alexander Volchek recorded in the car on the way from Silicon Valley to Santa Barbara: five to ten minutes before recording, an email came from OpenAI — within one to two weeks the company will merge ChatGPT with Work, and the chat with Work and Codex. Alexander calls this a substantial evolutionary shift for a system that today consists of ChatGPT, Codex, Work and Dots, not counting separate systems for researchers, biologists and education.
First he recalls how these products differ. ChatGPT is the basic chat that can already be given agentic work too: send an email, go through your finances, set up a weekly task. Codex, conceived as a system that works together with your computer, is seen by many as a tool for geeks and programmers, although in reality it is an upgrade of ChatGPT for recurring tasks, files, reports and presentations. The models and modes differ between them: in ChatGPT the medium mode may put on GPT-5.6 Sol and the maximum mode Astra, while Codex has Medium, High, Max and Ultra Code but no Pro, and within a single session Codex answered differently from ChatGPT. Work was conceived as an environment for documents, presentations and email without programming — in Alexander's view, a story copied from Anthropic's Cowork.
The teams that built these systems, Alexander says, didn't understand what they would achieve, and the leaders of OpenAI and Anthropic themselves do little work in Codex: "Where would Sam Altman find so much time to sit and figure out how to build various processes?" His example is an interview in which a blogger showed Greg Brockman an app for preparing to meet a guest, and Brockman called it a great system.
In response, Alexander dictates a prompt for Codex right in the car: assemble a dossier on Brockman as an HTML report — download his interviews from YouTube for the last seven years, transcribe them with a local Whisper, comb through no fewer than a thousand sources and social media, single out the trends, failures, right and wrong predictions, conflicts and his position on the twenty main questions about AI, then go through the author's own interviews and suggest new questions in his style. The work is to be broken into five to ten stages, starting with an audit, on GPT-6 Astra in Max mode. And he immediately adds: even this is already the old world of artificial intelligence, and the execution will have to be adjusted along the way.
The second half of the episode is about a mistake Alexander has seen since 2003, when he got hooked on data normalization and enterprise automation: ERP, CRM, various systems. Companies automated their existing processes without changing them — contractors went to the employees, and they wanted to make work easier for themselves rather than earn the company more. Great systems are built from scratch, like Airbnb and Uber, and old companies later catch up with the new business process. AI, he says, makes it possible to bring changes of a completely different class into a company and to build systems whose names haven't been invented yet, and he explains the split into Work, Cowork, Codex, Claude Code, Claude Design, Claude Duet and Dots by the narrowed perception of project managers who recreate processes already described on the market.
At the end Alexander returns to the news: OpenAI plans to remove the notion of unlimited work. Unlike Anthropic, ChatGPT kept good models essentially without limits, and the Pro plan was unlimited even on Astra; his cases of uploading tens of thousands of pages, a million calls and two hundred thousand tickets to analyze a company before joining it as a shareholder or on the board rested on that. Hundreds of thousands of startups solving a local task that AI will solve anyway, he says, have already died and more will die. The episode ends with a story about Santa Barbara and a preview of a conversation about Anthropic's IPO.
The episode reads the merger of ChatGPT with Work not as a rearrangement of tabs but as an admission that splitting into separate products was a mistake of a narrow view: the winner is the one who gives AI a big task in full and rebuilds processes around it, rather than speeding up the old ones.
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 44 segments: 0 identified, 0 mixed, 44 probable, and 0 unresolved.
Read transcript on a separate page
Loading…