Google Released Gemini 4 Argon: Is It Time to Leave ChatGPT and Claude?
Is it worth leaving ChatGPT and Claude for Gemini 4 Argon and Google Antigravity — and how close is the time when agents will predict our actions and do our work without instructions?
Ilnar Shafigullin explains that Codex already predicts the next step: when it comes back with an answer, it suggests your next prompt in grey text, and you just press Tab. Gemini 4 Argon, in his assessment, has joined Fable, Opus 5.5 and Astra in the same class about three months behind, but this is no leap like Gemini 3, which triggered a "code red" at OpenAI, and there's no point in running to Google. Alexander Volchek believes Argon was mainly released for shareholders, even though, according to the host, it is two to two and a half times cheaper than Fable and Astra, and Google Antigravity, to which Argon was added on October 13, has for him joined the ranks of Cursor, Codex and Claude Code. Ilnar relays Andrej Karpathy: even programmers with Pro subscriptions see a kindergarten compared to the labs. In Ilnar's test in China, ChatGPT as an interpreter lost to Google Translate, Dots has to be called, and in Alexander's projects AI itself sorts out and fixes requests. At the end Ilnar explains how to keep Codex on a server while the laptop is closed, and Ultra Fast, he says, is eight times faster for six times the price.
What to watch for
Key takeaways
Someone wrote on X that we may be just twelve months away from an agent that can watch your work and then do it without any instructions. The combination of screen recording, agent memory and a model running right on the device, he says, makes this possible already now.
At the time of recording Alexander still doesn't know whether he spent $37,700 on Amazon's servers. No bill has been issued: Amazon only shows the spending and suggests raising the limit to a hundred and twenty or a hundred and fifty thousand dollars. Whether Claude Code did it or Amazon is glitching is still unknown, and everyone is counting on a glitch.
Ilnar readily believes the forecast: people are fairly predictable. Already now, when Codex comes back with an answer, it writes in faint grey in the input field what you are most likely to ask for next, and all that's left is to press Tab and Enter. In straightforward tasks this is easy, and next, Ilnar believes, the story will spread to other areas of life.
According to Alexander, xAI claims five hundred million users on X, but by his data X's daily audience has fallen over five years from a little over two hundred million to supposedly a little over one hundred million. The five hundred million include Grok, and Grok, in the host's view, simply can't have three hundred million users.
Google Antigravity, according to Alexander, is not a new system: it launched on May 19, 2026, and it is not just a model but a shell for working with the model through a normal app, as in Codex or Claude. On October 13 Gemini 4 Argon was added to it, and for the host Antigravity has joined the lineup of Cursor, Codex and Claude Code.
Argon, in Ilnar's assessment, gained a little on benchmarks and brought Google back into the league but didn't turn it upside down. Gemini 3 once triggered a "code red" at OpenAI, whereas now Google, about three months behind, has reached the level of Fable, Opus 5.5 and Astra: most users won't notice the difference, and there's no point in running to Google.
To Alexander, Argon looks released for shareholders and the global trend more than anything else: much of its description is hard to verify. Yet, he says, it is about two to two and a half times cheaper than Fable and Astra, which cost roughly the same. He doesn't name prices per million input and output tokens: people don't understand them, and OpenAI's developers on stage gave no sensible explanation.
Tibo's limit-reset marathon goes on, and Alexander calls it "unprecedented generosity." Yesterday he burned one and a half billion tokens, the day before 2.1 billion, and before that five hundred million. One long task on 6.0 Astra in sales functionality, asked "Are you sure you're doing everything right?", admitted: "Oops, I went a bit in the wrong direction."
According to the description, Argon's overall capability index is the same as that of Fable 5.1 and Astra, and on Automation Bench, running work processes through tools, Argon supposedly does better than both. Yet the examples show the model drafting a corporate contract, whereas systems like Fable were said to hack into government systems, and Mythos, Alexander recalls, found ten thousand vulnerabilities.
According to Andrej Karpathy, as Ilnar relays him, even programmers with Pro subscriptions see a kindergarten compared to what is running in the labs, where swarms of thousands of agents are at work. Karpathy, Ilnar reminds us, now works at Anthropic, and Ilnar believes him more than he would believe Sam Altman: Karpathy hasn't been seen chasing attention and points.
People quickly get used to the level technology has reached. The computer was considered a calculator because it couldn't play chess, and when Deep Blue beat Garry Kasparov, people said it was just a fast calculator. The same goes for AI: what seemed cool in the days of GPT-3.5 is taken for granted today, which, Ilnar believes, is why Astra can seem worse than GPT-5.5.
Ilnar calls the results of testing ChatGPT as a simultaneous interpreter in China disgusting: Google Translate is much better. ChatGPT mixes up languages, saying "Ni hao" and then carrying on in Russian, and in Tatar it periodically switches to Russian and makes things up. With Google Translate, they reached agreement much faster.
In Dots, according to Alexander, you can't just press Voice: you have to call, and a real call goes through. Dots can't hang up by itself, and the whole voice conversation isn't saved in the chat, even though Alexander asked for the result to be sent there. The voice is decent, but the lack of thought is striking: models still handle talking and searching in parallel very poorly.
In Alexander's projects, requests are handled by artificial intelligence, Codex Astra, Fable or Opus 5.5 Max: it scans attachments with an antivirus, finds the real bugs, fixes them itself and notifies people via the Telegram bot and on the portal. Ideas go to the host for a decision, and a friend's spoken remark becomes a ticket with a thank-you for the request.
If Alexander can upload a small piece of code to his Roborock, such code can plant a virus there too. There is a type of virus that spreads among household appliances and leaves a micro-loophole: "a fridge, a TV," Ilnar suggests. And no one is responsible for holes in apps, and Amazon's contracts also disclaim responsibility.
For computations that run on a server for five days or more, Ilnar asks Codex to check the status every fifteen to thirty minutes and, within the allowed actions, restart whatever has crashed. So that it all doesn't stop when the laptop lid is closed, Codex or Claude is brought up on a server and connected over SSH, following the documentation on OpenAI's site.
Claude has five-hour limit windows, and an ordinary working day gets you two of them. An acquaintance of Ilnar's built a system that periodically sends Claude a minimal message like "Hi, are you there?": the window starts during the night, and in a single working day he uses the limit of three windows.
Ultra Fast, according to the hosts, speeds things up eight times for six times the price; the mode used to be only in 6 Astra and has now appeared in 6.1 Sol too. Alexander advises switching it on very selectively, when a system has gone down or you're stuck with support; otherwise it makes sense with a million-dollar budget or endless resets.
What this episode is about
The Sunday episode of ToTheMoon opens with a comment on X: we may be just twelve months away from an agent that will watch your work and do it without instructions, since screen recording, agent memory and an on-device model, according to the author, make it possible already now. Ilnar Shafigullin readily believes it: Codex already suggests your next prompt in grey text, and with one point two billion users OpenAI has, or can get, information on roughly every seventh person on the planet. Alexander Volchek answers with an example from life, car insurance that went up by six thousand dollars a year because of his newly licensed daughter, and says that the $37,700 Amazon story still hadn't been resolved at the time of recording.
The main topic is Gemini 4 Argon. In Alexander's view it is too early to compare it, but Google has Google Antigravity, a shell for working with the model like Codex or Claude, which, he says, launched on May 19 and to which Argon was added on October 13; for the host it has joined the ranks of Cursor, Codex and Claude Code. Ilnar takes a sober view of the newcomer: "it came out, so what" — Google, about three months behind, has caught up with Fable, Opus 5.5 and Astra, but this isn't Gemini 3, which triggered a "code red" at OpenAI, and there's no point in running to Google. Alexander believes Argon was released for shareholders more than anything else, even though, according to the host, it is two to two and a half times cheaper than Fable and Astra and won Automation Bench.
In between comes Tibo's marathon: Alexander burns one and a half, even 2.1 billion tokens a day, and after another reset the Codex widget on his iPhone shows ninety-nine again. Then comes Andrej Karpathy's tweet: the gap between those who understand AI and everyone else has become cosmic, and programmers with Pro subscriptions, in Karpathy's words as relayed by Ilnar, see a kindergarten compared to the labs, where swarms of thousands of agents are at work. Ilnar uses chess and Deep Blue to explain why we stop noticing progress, and Alexander explains why in the future his children won't learn to drive.
Next comes where AI is still dumb. ChatGPT as a simultaneous interpreter in China, Ilnar says, lost to Google Translate: it mixes up languages and makes things up in Tatar. Dots can't simply be switched on by voice: you have to call it, it can't hang up, and the conversation isn't saved in the chat. Feedback in big services, Alexander believes, doesn't work: to his fifteen messages Amazon only replied that it had passed them on. In his own projects he does it differently: AI sorts out requests, scanning files with an antivirus, fixing bugs itself and notifying people via the Telegram bot and on the portal, and a friend's spoken remark becomes a ticket.
Another topic is the Roborock robot vacuum with new lidars and bad software, into which Alexander wants to upload his own code, and the flip side of that wish: viruses in household appliances, trashy apps that Apple can't cope with, and contracts under which services aren't responsible for anything. At the end Ilnar explains how to ask Codex to check a long process every fifteen to thirty minutes and how to keep Codex on a server over SSH while the laptop is closed, and an acquaintance of his uses minimal messages to get three five-hour Claude windows in a working day. Ultra Fast, according to the hosts, is eight times faster for six times the price, and Alexander advises switching it on only for something super-urgent.
The episode answers the question in its title without hype: Gemini 4 Argon, in Ilnar's assessment, has merely caught up with Fable, Opus 5.5 and Astra, so there is no point in leaving ChatGPT and Claude for it, and Google Antigravity is more a reason to try than to move. The figures — Argon's price two to two and a half times lower, the dates of Antigravity's launch and Argon's arrival, X's and Grok's audiences, billions of tokens — come from the hosts and need checking, Karpathy's words are relayed second-hand, and the $37,700 story was unresolved at the time of recording. The most practical parts are Codex on a server while the laptop is closed, a poller for long tasks and AI that sorts out requests and fixes bugs by itself.
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 154 segments: 154 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…