Stop Just Asking ChatGPT: 4 AIs That Can Do the Work for You
Why would an ordinary person need Codex, Claude Code, Cursor or Google Antigravity if they already have ChatGPT — and how can you tell which of these systems really does the work for you?
Codex took several months of bookings from Alexander Volchek's email — Expedia, Booking, Chase Travel, Amex Travel, Marriott — found the hotels that had not applied breakfast credits, wrote to them itself and kept up the correspondence until some of the hotels refunded the money. It also helped reassign a speeding fine in France to a friend and followed the request until a decision was made, while Dots tracked a prepaid envelope with documents all the way to delivery. Codex connected by itself, through the browser, to the Amazon cloud where Alexander's cameras store video without any API. The task today, according to the host, is not prompt engineering and not settings like "slash goal" in Antigravity or long-running agents in Cursor, but a goal: "Analyze my emails to see whether I've been charged extra money." Cursor and Antigravity, by his estimate, lag behind Codex and Claude Code because of model infrastructure and memory; for development he chooses Claude Code with Opus 5.5 and Fable, for everything else in life the combination of ChatGPT, Codex and Dots, and he advises keeping two subscriptions — for $100–200 and for $20. A live test in the episode also showed the limits: Dots came up with a sum of two hundred dollars at Alexander's favorite cafes, and he immediately saw that the system had not seen all the accounts. He calls seven parameters decisive — from the ability to work for days without losing context to who decides that a task is finished.
What to watch for
Key takeaways
Alexander considers one of the main misunderstandings about artificial intelligence to be the question of why someone who already has ChatGPT needs Codex or Claude Code. He compares four systems — Codex, Claude Code, Cursor and Google Antigravity — not by programming and technical tests but by what they already do in ordinary life, professional work and business.
ChatGPT has become a generic name, the way Pampers or Xerox once did, and most people perceive modern AI as ChatGPT. The companies themselves add to the confusion: Anthropic "killed" Cowork and is phasing out Claude Design, and a switch between ChatGPT and Codex has appeared in the ChatGPT app for Mac.
Within ten years the understanding of who can consider themselves a programmer will change. Alexander graduated from university in 2005, then built systems through the hands of teams, and now he can build them himself: the notion of efficiency has changed, and in some situations doing it himself has become more advantageous.
Codex and Claude Code are joined by Google Antigravity with the just-released Argon model and by Cursor, which Alexander suggests calling Cursor xAI or Cursor Grok: the name itself, in his view, will disappear and dissolve.
Within the next year, applying automation in your own life will become as simple as registering an email account once became, and it will become an ordinary part of life, like paying for electricity or internet. That is why it is worth getting going a bit earlier and trying to do something serious.
Email and services are easiest to connect to Codex: behind it stands ChatGPT, which for Alexander surpasses Anthropic's system. But as a development system Claude Code is, for him, "cooler than Codex": Opus 5.5 and Fable in Max mode are better than 6 Astra and 6.1 Sol.
The simplest task is to connect your email to Codex or Claude Code — not to clean out the inbox, but so that the system goes through it every day and the person interacts with it without opening the mailbox. In the same way you can give access to Telegram, photos, files and bills.
Codex can be integrated with banks, but so that it can't spend money: read-only or with a code to enter; or at least upload statements once a week. The new Finance module in ChatGPT is integrated with a large number of banks.
After a trip around Europe, Codex took the Expedia, Booking, Chase Travel, Amex Travel and Marriott bookings from the email and found the hotels that had not applied breakfast credits — including in New York and Hawaii. It wrote to them itself and kept up the correspondence until some of the hotels refunded the money.
More than twenty of Alexander's cameras store video in the Amazon cloud without any API, and previously third-party "middlemen" were needed. Now Codex has connected to these systems by itself and downloads the videos: integration without API and MCP is becoming less and less of a problem.
Codex helped reassign a speeding fine in France to a friend and kept track of the request until the French system had reviewed it, and Dots reported on its own that a prepaid envelope with documents was in transit and would arrive the next day.
"Slash goal" in Google Antigravity, long-running agents in Cursor, agents.md in Codex — that's a billion settings, as in Windows in the nineties, and AI itself will do the setup better. The task today is a goal: "Analyze my emails to see whether I've been charged extra money"; if the system doesn't ask or produces "complete gibberish," it doesn't work for this task.
Appointment-booking modules, calendar apps and payment "middlemen" with ads will become unnecessary: your own appointment-booking module is now made "with a snap of your fingers," and a seller needs the buyer to stay on his website.
Google Antigravity and Cursor lag behind Codex and Claude Code because of Anthropic's and OpenAI's model infrastructure. For Alexander, memory is decisive too: over thousands of chats, the combination of ChatGPT, Codex and Dots already knows where he lives and what he does.
Alexander advises having two systems — for $100–200 and for about $20 — and testing on hundreds of different tasks: even the most expensive models, such as ChatGPT-6-Astra, may fail to finish an elementary task and go off track.
Codex sorted tens of gigabytes of iPhone voice memos spanning fifteen or sixteen years into folders, spinning up a cheap local model for speech recognition. Through Dots you can ask it to find something from four years ago — in your email, photos, files, audio recordings and notes.
Alexander puts the probability of Apple disappearing at tens of percent, and for Microsoft much higher. He supports Elon Musk's Macrohard project and believes that probably ninety percent of solutions can be rewritten now.
Right during the recording, Dots found eighteen payments on fifteen days at Cafe Venetia and five on four days at Satura. Alexander immediately saw a mistake: a sum of two hundred dollars is "complete nonsense"; the system lacked access to all the accounts.
The first parameter of comparison is the ability to keep going on a big task and not stop after an answer. Alexander has spent 63.3 billion tokens in ChatGPT, and on peak days in July the system was eating as much as five and a half billion a day.
Dots often comes back an hour or a day later with a new conclusion. Alexander believes a task should never be finished: real artificial intelligence constantly interacts with a person, and he would pay two to four thousand dollars a month for such a system.
What this episode is about
A solo ToTheMoon episode in which Alexander Volchek takes apart what he considers one of the main misunderstandings about artificial intelligence: why someone who already has ChatGPT would need Codex or Claude Code, especially if they are not a programmer. ChatGPT has become a generic name, like Pampers or Xerox, and the companies themselves muddle their products: Anthropic "killed" Cowork and is moving Claude Design inside Claude Code, while OpenAI's Mac app has a switch between ChatGPT and Codex. The host compares four systems — Codex, Claude Code, Cursor, which after the SpaceX purchase is more properly called Cursor xAI, and Google Antigravity with the new Argon model — not by technical tests but by what they already do in life, professional work and business. He is sure that within the next year automation will become as simple as getting an email account once became, and an ordinary part of life, like paying for electricity.
Alexander calls himself a super-professional user of Codex and Claude Code: it is more convenient to connect your life to Codex, which has ChatGPT behind it, while development he entrusts to Claude Code with Opus 5.5 and Fable in Max mode. The first task for an ordinary person is to connect email — not to clean out the inbox but as a new interface between yourself and dozens of systems: Telegram, photos, files, bills, utility payments, banks with read-only access, the Finance module in ChatGPT. The episode's main case is a trip around Europe: Codex retroactively found, in Expedia, Booking, Chase Travel, Amex Travel and Marriott bookings, the hotels that had not applied breakfast credits, corresponded with them itself, and some of the hotels refunded the money. It also helped reassign a fine in France and followed it until a decision was made, while Dots tracked a prepaid envelope all the way to delivery.
The host compares complex settings — "slash goal" in Antigravity, long-running agents in Cursor, agents.md files in Codex — to Windows in the nineties: AI itself will do the setup better, and a person needs a goal, like "Do it!" for a tax return. If the system doesn't ask questions or produces "complete gibberish," it doesn't work for this task. Hence the forecast about services that will die out: appointment-booking modules, calendar apps, payment "middlemen" with ads; even Booking, Airbnb and Stripe will get clones. Cursor and Antigravity, in Alexander's view, lag behind Codex and Claude Code because of model infrastructure and memory of thousands of chats; he advises keeping two subscriptions — for $100–200 and for $20 — and testing systems on hundreds of tasks, not on one.
The episode also has personal cases: Codex sorted fifteen or sixteen years of iPhone voice memos into folders, spinning up a local model for speech recognition, and Dots serves as an easy entry point to both Codex and ChatGPT. Alexander calls Microsoft "simply a dying infrastructure," supports Elon Musk's Macrohard and believes that probably ninety percent of solutions can be rewritten now. Right during the recording he asks Dots at which cafe he spent more and immediately sees a mistake: the system lacked access to all the accounts.
At the end the host names the parameters by which the four systems are worth comparing: the ability to work for a long time — he himself has spent 63.3 billion tokens in ChatGPT — behavior when the history stops fitting into the context, memory of what you have taught the system, coherence of parallel agents, the decision on when a task is finished, availability from all devices and handling independent tasks at the same time. In his view, a task should never be finished: real artificial intelligence will constantly interact with a person, as in the film "Her."
The episode is useful as a practical guide for those who so far use only the chat: real cases show how an agentic system differs from ChatGPT — it collects data from email and banks by itself, writes emails and sees things through to the result — and the parameters for comparing Codex, Claude Code, Cursor and Antigravity are named. These are the assessments of one experienced user who speaks openly about his preferences: ChatGPT and Codex for life, Claude Code for development. The figures and events — the SpaceX–Cursor deal, the host's 63.3 billion tokens, the probability of Microsoft and Apple disappearing, the PHP bug — are given as he states them and need checking, and the live test with Dots showed that even a strong system makes mistakes if it doesn't have access to all the data.
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 155 segments: 155 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…