Hi everyone! You're on ToTheMoon: tech insights and news from Silicon Valley and around the world. And today I want to record a special episode that is probably more about Claude Duet. It's a direct competitor to ChatGPT Dots, but in reality I want to pay attention to four systems that are now regularly compared as similar. These are Grok Bot, Facebook Muse, Claude Duet and ChatGPT Dots. And overall, I'll try to show you what can be compared between them and what can't. The thing is, when Claude Duet came out, and that literally happened a week ago, they announced that it was launching in a kind of pre-release mode.
That is, they'll roll it out only to Pro users, only in the US and only through a waitlist. And it took me probably four or five days to get access through the waitlist, and the internet exploded with claims that ChatGPT Dots was dead. And that a super cool system had appeared. And it seems to me this is the hype market that exists: for two or three years everyone has been predicting that ChatGPT will die. Or they keep writing that some new model has appeared that someone built and that is a thousand times cheaper than OpenAI's models or Anthropic's models.
And, accordingly, that's it. Death to artificial intelligence and death to these systems. Or Elon Musk speaks and says that the market of companies that provide compute and infrastructure has changed. And the internet is bursting with headlines saying that artificial intelligence, the AI market, is a bubble and it has vanished. Although he didn't say that. By the way, in that statement he said that companies that generate no revenue other than from infrastructure will have certain problems.
But in particular, by the way, this concerns a company I like quite a lot, Nebius. Yes, it's the company of the former owner of Yandex, or a group of those people. And since they provide infrastructure to companies in artificial intelligence, for using different AI models, they are, accordingly, a provider of that infrastructure. So, as an example, they will have a problem, for instance. And the market really is shaky, by the way: the AI and tech market over the last ten days, roughly over the last two weeks, has seriously blown up.
A number of companies have really, seriously blown up. But it's always important to look constructively at what is happening in reality. And for you, as an ordinary user, not to make a mistake in your choice, in your preferences, or in actually starting to use something. People still don't properly start using Codex.
And one of the problems, by the way, why people don't properly use Codex or Claude Code is that, in essence, a person doesn't know what to do in it. Here on the channel we try to give a huge number of examples. All the episodes around this video are very much about different examples of where you can apply Codex or Claude Code in your life. And the same with Dots. What I see is that people don't yet fully grasp what Dots is, because they're used to basically opening a chat in ChatGPT, asking some question, finding some information.
And they don't understand the foundation of this system at all, or how this system will develop. We had a huge episode about Dots. And here, of course, is a continuation about such systems. Let me remind you what these systems even are, such as Duet or ChatGPT Dots or Grok Bot, or the overall announced story of Muse at Facebook. I didn't drag Muse in here for nothing, because a lot is written about it. Although, of course, in terms of the quality of that system it's a completely different level. But still, what is this system?
It's a system that gives you the ability to work within an interface that is probably similar to a messenger. It's still not a chat interface like ChatGPT. Yes, it's much simpler. There are no tasks, there are no projects, there is no base. You don't see that list. You open the app, launch the screen, so to speak, a screen. We'll sort out the apps in a moment, because Dots, for example, has no app, while Claude Duet has its own separate, dedicated app. Yes. And that, by the way, is a nice touch.
I think it's a nice touch. And ChatGPT should probably release a separate app for Dots, but that, again, is a question of interface. We'll discuss it in a moment, okay? Because ChatGPT already has a huge number of different entry points, and Anthropic often gets it wrong. They released a separate app, for example, for Claude Designs. Now they're shutting it down, and it won't exist anymore. It will live inside the shared infrastructure. Or Anthropic, for example, had Cowork and a separate chat; at the beginning of September, on September 16 to be precise, they merged Cowork and chat, and that's gone now.
And of course people get confused. But what's the point? You have essentially one window that you write into. And which then relays data to you. In fact, the idea itself is unbelievably cool. And Dots really is limited in availability. Right now it's available only on the Pro plan and only in the market, basically the US market. And around the world, if you look, it's appearing in some countries. I periodically see that people have gotten it. But even now a friend of mine flew in to see me, and if he connects his Asian mobile SIM, he has a Singapore SIM, Dots disappears for him right away.
As soon as he connects to my Wi-Fi, for example, Dots appears for him right away. And that's a problem. Well, it's a really serious problem. Although my sister, at the moment Dots opened, didn't get Dots at all on the Pro plan, even in America. There's no Dots in Europe, and it doesn't even seem to be promised, since it's a big problem in terms of complying with European Union rules. It's a real problem of complying with European Union rules. This window is very interesting because it essentially simplifies your life a lot.
That is, in one window you interact with all of OpenAI's infrastructure or all of Anthropic's infrastructure. You're supposed to interact. We'll figure out in a moment how it works. Or you interact with all of Grok's infrastructure, xAI. Or… It's Elon Musk's company, a very serious competitor for OpenAI and for Anthropic. A serious competitor, because it's a huge, incredibly serious resource in terms of infrastructure and money. By the way, Elon Musk said that SpaceX will be worth a hundred trillion dollars. And I want to point out to you that this is a very serious statement in terms of the market and in terms of his approaches overall.
Someone said the company might be worth twenty or thirty trillion. He said: "I think it could be worth a hundred." Some of you will say something, and it will be interesting to hear your opinion here. And by the way, write in parallel which of the systems listed above you use, or who wants to use them, or who is waiting for them to open, or who understands the difference. What do you think about this too? It's great to see your thoughts. So, someone will say that Elon Musk talking about a hundred trillion is pure hype.
I want to say, of course, that clearly, on the one hand, it's a marketing kind of story of the right approaches of companies that are publicly traded, and of share prices. But on the other hand, you understand, any such statement is later remembered and examined: who's talking empty words and who isn't. And still, the appearance of such numbers in the world, reasoning about such numbers, is fundamentally interesting. I'll remind you, by the way, that I have said many times that Anthropic and OpenAI should, in essence, be worth more than ten trillion dollars.
And that OpenAI should be worth more than ten trillion dollars has always been somehow obvious to me, that a serious provider of artificial intelligence, of the current systems, over a horizon of, I don't know, five years or three years or ten years, should in principle be worth that much. That's an adequate reality of what's happening. Or there would have to be incredibly large competition, there would have to be a great many such companies, and… it's unclear how that would happen.
So, well, that's a separate topic, a separate discussion with you. So, the single window. What is it? They're now testing the single-window functionality through one chat. That is, you can write into one chat in Claude Duet. This chat creates different tasks by itself. This chat comes to you on its own and reminds you about your various daily tasks. This one chat personally works in different project systems by itself. That is, one separate chat. And you can apply this to Dots too.
Partly this applies to GrokBot and Muse. They are still systems of a certain next generation; they don't have such cool infrastructure. I recently compared systems, Claude Code, for example, Codex, and for Grok it's Cursor. There was also one more operator, by the way, Google Antigravity, let me remind you. And given that Argon, a new system, has appeared, Google Antigravity is still becoming a fairly serious player in the market. By the way, what does everyone think now about Google Antigravity and Google's new models? That is, how much do you think they are at Fable's level?
I still think that Anthropic's Fable level is a kind of top tier of models. I don't count OpenAI's Astra there yet, although I'm a huge fan of OpenAI and ChatGPT, I don't count Astra there yet. Look, here's a certain problem. When you have one window, it's very important for everyone to understand that this isn't simply an agent-based way of working. When you have one chat, or you have twenty different sessions in which you individually created agents, or you have one chat in which you manage, among other things, different sessions.
I'll note that this capability already existed. Right? Here's an important point for all of you: these are different approaches to managing models in general.
There are people who use models in some really old-fashioned way, from when, a year and a half ago, someone would say: "I manage fifty agents in parallel. One agent does this for me, another does that." And people started setting things up so that one agent, I don't know, finds news, a second agent updates information, a third agent does design, a fourth agent does layout, a fifth agent creates a report, a sixth agent sends out sales information, and so on. In today's reality, in the modern reality, this is absolute nonsense. It makes sense for you to write out the whole task completely in one session, if it's within one project, and one session will carry it out.
Will it use different agents? It may use hundreds, thousands, dozens of agents, or not use agents at all. That shouldn't concern you. What difference does it make to you how many agents are working there? Statements even from programmers at serious companies like OpenAI and Anthropic that they simultaneously use, that a thousand or two thousand agents are working for them, are absurd in themselves. It's a kind of marketing statement that's fun for everyone to hear, that different agents exist. Now we're seeing that ChatGPT Dots has appeared.
If you simply call ChatGPT Dots an agent, then this agent is cooler than all those thousands of agents that came before. In fact, ChatGPT Dots or Claude Duet will run different tasks, different sessions in parallel, and will use Anthropic's and OpenAI's infrastructure each in their own way. Completely differently, by the way. We'll get to that question in a moment. And carry out a huge number of different tasks. And forget the word "agent" here altogether when it comes to this approach.
And notice that now, when you set new tasks, the question is often not the quality of writing the prompt, but your ability in general to adjust this task a little in terms of your overall approach. How do you want it to be implemented? Because, for example, when you say make a page design, it's not a question of prompt quality but of model quality, plus your own vision of how it could be. This vision is an incredibly important thing. That is, are you able to imagine in advance when creating, for example, some page, some platform, or your website, or some report format, what you want to get?
It's one thing when the model imagines for you how it should be. It's another thing whether you understand to some extent how it could be presented at all. Because the model can give you very different options. That doesn't mean some option is vastly worse or vastly better, but they will be very different. While preparing for today's episode, I spent a very long time asking ChatGPT Pro to compare these four, let's call them, bot systems. Or these new systems of a new approach. It's even hard to call them by any word. It's unclear what word to call them at all.
Well, some call them a permanent assistant, some call them some kind of agent system. In fact it's unclear what word, what word to use anymore. And because with ChatGPT it's clear, it's all this… If we listen to the whole world, everyone will say: "ChatGPT, ChatGPT do it, ChatGPT do it." It's become a generic name, right? And I said in some recent episode that it's like Xerox or like Pampers. That is, such a well-established name, it's clear what it is. Everyone immediately understands what's meant.
Although there may be many companies that make it, and different systems may be used. Of course, because of this people get incredibly confused. It's one thing when we compare diapers, Pampers and Huggies, and the comparison there is absolutely minimal. It's another thing when we start comparing GPT-like systems.
And here the point is that they start to differ from each other enormously. And these four systems, like Grok Bot, Claude Duet, ChatGPT Dots, or Facebook's Muse, Meta's, yes, Meta's, Meta Muse, they differ from each other incredibly strongly, simply fundamentally, once again. And there's an important structure here. That is, when you do open a session in a chat and tell it to do something, this session has, in terms of the AI model, a bunch of certain limitations. First, there's the notion of a context window. So, Grok Bot, for example, or ChatGPT Dots, or Claude Duet essentially have no context window.
Look, it isn't there. Moreover, if we look at Dots, unlike Claude Duet, by the way, Claude Duet made a separate app, which is a separate entry point. Not bad, by the way. A neat thing, because every time I open ChatGPT it's a whole problem. I opened ChatGPT, I need to go to Codex. I opened ChatGPT, I need to open Dots, or I opened a ChatGPT chat. It's a real, constant problem. The ChatGPT Mac app now also sees Codex chats. Codex sees only its own chats. Separately there's Dots, and the web version of ChatGPT doesn't see Codex chats. You have to switch to Codex. Such a very, very clunky story.
They work differently everywhere. Moreover, for example, in the web version of ChatGPT, chats inside projects aren't shown as recently used, right? While in the Mac app, I think, they are shown. That is, such very, well, clunky interface management. They're searching for it, they'll find it. And I want to tell you that I'm sure such systems, such an entry point as Claude Dots, as ChatGPT Dots or Claude Duet, well, the question here, it's important to understand what the company is focused on.
ChatGPT is still focused on the mass market. Anthropic is relatively focused on the mass market for now. So, and ChatGPT is still more than a billion users, it's, well, a serious global leader. So, what is a single entry point? A single entry point can be a chat, it can be you calling somewhere by voice. That's exactly why they separately built into Dots the ability to call somewhere. And this, in principle, is a speaker that stands there and answers you somewhere. It's the same Dots, right? It's just some visual, something displayed somewhere. It's the same Dots. And Dots can be everywhere: in the car and anywhere at all.
And by the way, probably the first innovator in this, in this movement may have been Grok, right? Maybe Facebook was too. But Facebook strongly, Meta loses very strongly in such systems. Perhaps they could also be called an innovator, because Meta AI was basically a kind of add-on and still is, in Facebook, for example, or in WhatsApp. When you open a single window, conditionally without chats, and interact centrally in it. But Meta AI, obviously, didn't create different tasks and different sessions from the start.
But in reality it's an enormous, big infrastructure of life, it's a different word. It's just that locally you can say it's an operating system, to understand what it is. When you have a phone and OpenAI is basically installed on it. And further inside there are all the additional add-ons and all the connections. You call from OpenAI, you send messages from OpenAI, you use different channels from OpenAI, messengers, whatever, you open browsers and everything else. You do it from OpenAI wherever it's needed. Clearly, behavior, behavioral scenarios will change very seriously; many of them will become completely different, not the ones we're used to using.
Even if we talk about the scenario of interacting with email clients, it's important to understand that the email client itself, for example, dies at some point. Right now I've started sending 80% of my emails through Dots, or rather through OpenAI's infrastructure, OpenAI and ChatGPT. And there's an interesting aspect here: I still do send email. But then use cases are born where you stop using email, because. That is, you start interacting with other structures. As I gave you the example: if my wife turns on Dots, she hasn't turned it on, she has a Plus subscription, right?
And maybe she would already be living in it. Or if they open the ability for Dots to communicate with each other, I'd get her a Pro subscription, or she might turn it on herself, so that we could message each other. What does that mean? That our Dots will start interacting with each other. That's an unbelievably cool thing. Of course, a huge amount of communication over messengers disappears, a huge amount of communication via email simply disappears. Before, something would come from school, and my wife would forward those emails to me.
Now my Dots sorts through all the emails that concern school and sends her messages itself through iMessage. If not, if Dots is able to compete, it will send them to her own Dots. Further, essentially, Dots now reads email, the schools' email. But in reality schools have platforms, and it's much simpler for Dots to connect to those platforms, to be able, for example, to communicate further through those platforms, including on behalf of the schools, or to take certain actions that can't be done via email.
Why do I need email then? It will connect directly to these systems. That's essentially what happened with banks. OpenAI is integrated with financial banks, very seriously integrated, through thousands, I think ten thousand, financial institutions. And clearly this isn't available in many countries. And understand, someone says: "Alexander, you're talking about some systems that aren't available to us, or that we can't open, or we don't have such a plan." I understand, but you see, comparing and offering me DeepSeek instead is pointless, because DeepSeek doesn't have what OpenAI has.
And when I tell you about many things, I show you certain use cases that you, for example, can expect for yourself. Or you'll know that such a use case already exists. You won't be under the illusion that you're, for example, using AI systems in a cool way by using, say, a French model, or an Arab model, or a Chinese model, or a Russian model. You won't be under an illusion; you'll understand that these models don't have this, they don't have this level of systems, this level of infrastructure. It simply doesn't exist.
And of course certain things exist, and certain individual tests may be better, but right now we're talking about a whole big package. And so, when my Dots is integrated through ChatGPT with all my financial instruments, what happens? I already open the bank's mobile app less often. I no longer get alerts or other messages through email, because my Dots is integrated directly with the systems. There's a nuance there. For example, Dots won't see, I don't know, statements, certain statements; it won't see letters from the bank, for example; I don't know whether it sees them or not. Probably not.
Or maybe there's some delay, for example, in receiving information. For example, a delay with the credit score report. In the US there's this thing called a credit bureau, and they give you your credit score. I see that ChatGPT seems to have the information a bit later than the actual score from that credit bureau. Although that's not certain. Perhaps they have exactly the accurate, verified information. And what the mobile app often shows isn't fully real information. But it turns out I've left email as the middle layer, or left SMS as the middle layer. That is, the use case has changed.
Systems like Dots or Claude Duet change things, or ones like GrokBot, yes, and I recommend them. GrokBot is more accessible now than Dots and than Duet. It's openly available, it's open, many people can sign up for it, just like for Muse.
It's important to understand that these systems change your whole approach to interacting with artificial intelligence in general. If you work with artificial intelligence a lot, then once you start using it, you'll of course be interested in connecting a system like ChatGPT Dots or Claude Duet. You'll start seeing different, well, clunky sides of these systems. By the way, I liked Claude Duet much less than ChatGPT Dots. Perhaps because Claude Duet is very fresh, right? Let me remind you about Dots: I also had questions about it in the first weeks.
It couldn't work with other chats that had already been created. It saw certain different sessions poorly, it couldn't launch inside sessions, it was hard to call Dots from sessions, and so on. In general, I think the presentation OpenAI made about Dots, or the use cases they send out, are kind of unreal, not true to life, very narrow, sort of semi-marketing. And in general OpenAI has a certain problem with informing customers about what they can do in different systems. Just a couple of days ago I got a newsletter again.
I terribly don't, this, this, this is the quality of one in ten of the newsletters OpenAI sends its customers all the time, well, not all the time, in any case, yes. They sent me how I can use Dots. That is, you can see this newsletter is detached from reality, from how I use it every day and what different use cases I run through it, what I test, what I check there, what I look at and so on, yes. By the way, why aren't Dots and Claude Duet rolled out to everyone? Because it's a hard, hard infrastructure task and very expensive for these companies.
Let me remind you that if you buy a paid plan in ChatGPT or in Anthropic and use it to the max, burn tokens to the max, and in ChatGPT there isn't even, in essence, a concept of a maximum. Because ChatGPT, the chat itself, is essentially unlimited on paid plans; it depends on the model and so on, yes. There, ChatGPT for five hundred dollars is unlimited everywhere, for example, and for two hundred dollars it's almost unlimited. Some systems are limited in terms of, for example, Astra or certain session windows.
For twenty dollars there's a different one, or for thirty dollars there's its own different unlimited, and so on. And in these unlimited plans, by the way, in everything else, you really, really have to dig into it. But the point is that if you buy a paid system and use it to the max, then usually you are a very serious loss for these… well, that is, for these companies it's a very serious loss, just an incredibly heavy loss in terms of the cost of these tokens, if you were buying them separately, yes.
So, some people like to say: "Oh, well, you don't know how much tokens exceed the real cost." By the way, what does everyone here think about this? Opinions always differ a lot. And, well, look at the scoreboard of these companies' revenue and expenses, and you'll see that these companies are loss-making, extremely loss-making, yes, very heavily loss-making. But it's important to understand that a small number of customers in the world use the load, the infrastructure, to the max.
That is, if everyone used it to the max, it wouldn't just be loss-making, it would be a terribly loss-making organization. We simply don't use this infrastructure to the max. It's a matter of time. And, by the way, this is one of the fundamentally serious challenges. That is, it's important to understand that if right now the whole billion users even bought the twenty-dollar or the two-hundred-dollar version and started using it to the max, the company would start bearing trillion-dollar costs.
They, by the way, wouldn't be able to serve it. Well, imagine a billion people starting to put such a serious load of billions of tokens a month, for example, even if they bought the five-hundred-dollar version or the two-hundred-dollar version. That's a very serious infrastructure load. So, of course, Dots isn't being rolled out in many places. And when you start working with these systems, you'll see that they have various limitations. For example, why didn't I like Claude Duet?
Because it can't work with the sessions I already had in Claude. In general, I think this is the first time Anthropic has released such an anticipated product so badly. I hope this problem gets solved, but it's a really huge problem. That is, as a regular user, I was given access to this product. But obviously, those who have the expensive versions, the Max versions, at two hundred dollars, roughly, people who use these systems more often. And obviously I have a lot of chats in Claude. I immediately opened Duet yesterday when I got access, and right away started telling it: "Listen, in these tasks change this, in these that, in these that, in these pull it together."
And it immediately sort of started reading these tasks. It saw my other tasks, but it couldn't do anything in them. And it says the following: "I can only work in tasks and sessions that I create myself, that I will manage myself." And that doesn't interest me, because I already have infrastructure built. And what's more, I want to say that I'm worried. What will happen if you create sessions? And who else will be able to work with these sessions? So it turns out you've got such a useless thing born, for now. They announce various stories.
For example, they announce that it's a system capable of working with your email and correspondence. And you go: well, all the chats were already capable of working with email and correspondence. Why do I need a central window today? But the central window is actually cool, yes. You sort of process all the chats from one window, process all the emails, but then again, you could process all the emails from a chat, right? And you know, you have to get used to this mode. Whoever has used Dots: you write, like, "Listen, and…" It sends you, for example, that you've received requests for such-and-such a contract.
Well, by email, for example. I write to it: "Study the contracts for me, draft replies." And at the same time I see that above it wrote me information that some system sent me that I'll have a big charge, I need to sort it out. And above that it also sent me a reminder that I have some flight, and I need to prepare for this flight. And at the same time I'm walking down the street and solving some question with it, for example, about what I need to find, how to do something on my property, yes?
Yesterday we were walking in San Francisco with Max and Dima and discussing with Max printing on a 3D printer. And I was solving everything through Dots. And you've got a mess in Dots. You have to understand that it becomes a mess. It's one window, one window of correspondence, and in this one window of correspondence it becomes a mess. That's the problem with all single-window systems in general. And it's like Slack or project management through Telegram. A huge amount of management… That's a problem.
That is, when you get a lot of tasks and projects, and if some message got lost, that is, your Dots or Claude Duet forgot about this message or didn't remind you, you've got a problem. I want to say, by the way, that ChatGPT Dots doesn't remind you about many things. That is, look, what makes this system incredibly interesting, why is it expensive, why isn't it rolled out to everyone? Because it's a system that is constantly listening to the context of your interaction with OpenAI, with different chats, with different sessions, sessions in which you make some reports, or sessions in which you solve some tasks, or sessions where you told it to listen to your email, for example, to do, or to send some messages somewhere, or something else.
That is, this system is essentially meant to be in constant interaction with your various infrastructure. But right now it listens to this infrastructure only very partially. That is, what's more, if the memory in your infrastructure was updated, you used it somewhere in a chat, your memory got updated, then supposedly Dots should know about it. It should know that a change happened. Or you launched something new in Codex. It should, it should know that this change happened. And that's a very expensive thing.
That is, being in everything all the time is very expensive. And of course it doesn't listen to everything. It might send you once: "I'm reminding you that you, roughly, well, you, I don't know, got an email, and tomorrow you'll be charged five hundred dollars if you don't take a certain action," and never mention it again. That is, if they don't send me a new email, it stops talking about it. And that, of course, is a big, serious problem. What can we see? That both Claude Duet and ChatGPT Dots are still very heavily tuned to their own personal context window. There, in essence, the context window doesn't exist, right?
Because this huge correspondence is, on the one hand, a context window, and on the other hand, a huge number of different context windows. That is, it's a fairly unstructured memory. And in general it's important to understand that with modern models it's not a matter of working, of working super hard, with the context window. Today there are already a huge number of statements from model developers, among others, that you don't always need to compact the context window. Let me remind you, well, what does compacting the context window mean? For example, my longest task in Codex to date, that I've ever run, I'll tell you, in terms of days, is fifty-eight days.
And the longest task in terms of continuous running time is eighty-eight hours. How can you check this, by the way? Go into Codex, click on your profile and look. Write in the comments, by the way. It's interesting who has how much. Right now I have 64.7 billion tokens. Peak token usage was 5.4 billion. The longest task, eighty-eight hours, ran for me forty-seven minutes. The longest work in terms of time, it shows days there, yes, that's the last one, you can see it in days, fifty-eight days.
That is, how much, and there's the current maximum task that I use, mine, and how much the maximum has ever been. And for me it turns out to be one and the same, right? Because it's been running for fifty-eight days. And at the DevDay conference, OpenAI, a month ago, I was right here in San Francisco. There, one of the heads of development was saying that he has sessions that run for two months. Well, with fifty-eight days I've just about reached that kind of session. Although the systems, both ChatGPT and Claude, can still suggest shrinking the context window to save resources.
And that, by the way, is a big question: whether you should shrink it, whether quality will get better or not. I have some tasks in Claude Code that translate text into different languages, because on my site, even if you look at my personal site,
well, first, I have different tasks not only on my site, but here on besolid.com, where there's a separate big infrastructure, the ToTheMoon site, translation into several languages is done there. So, I'm recording this episode in Russian right now. This episode will appear on the site. And recently Claude suggested I use a cheaper model for translating text. But what's interesting is that, on the one hand, I can save tokens, and on the other hand, I can lose on translation quality. Someone will say: "Oh, well, translation is an easily solved task."
But here's the question: what vocabulary is used, what meaning did the author want to get across? And that's already a question of the model's quality. How much this model double-checks it. Okay, then you can translate, for example, with a cheaper model, then process the text, the text, for example, with a more expensive model. And the question is: if we had translated and processed the text with the more expensive model right away, would the model have spent more money overall or less?
Look, as of today, this answer basically doesn't exist. We know, for example, that the Fable model is definitely more expensive, 5.5, definitely more expensive than the Opus 5.5 model. At the same time… Fable 5.1, yes, is more expensive than Opus 5.5 at Anthropic. At the same time, you and I know that every model has an operating mode. And there's a mode, it's one thing when you use low mode, low quality in terms of the model's effort, and there's max mode, for example. And there's even Ultracode mode.
That's when the system runs in, well, super, super usage mode, and tokens get eaten there at different quality. So, the story is that Fable 5.1 is, I don't know, five or ten times more expensive than Opus 5.5. At the same time, the quality of the result, or the number of tokens spent, by the way, can be lower for Fable than for Opus. It all depends on the task being solved. You can give examples now of who had what and where, yes. But if we start figuring out a huge number of different examples, we'll see that everything depends very heavily on the task.
And if you're not a specialized employee, one in ten thousand at a company, developing one, or at least one in five hundred or a hundred, developing some solution inside, but are building these systems individually or with a small team, or building them for marketing, or for sales, or for design, or for analytics, or for finance, or even for yourself, then this choice of which model you use is, well, just like looking for a needle in a haystack. That is, which model should I set to solve this task? Fable 5.1 in max, in Ultracode, in low, in medium, or Opus 5.5? What will be enough to solve this task?
That is, if you bought a plan, for example, for a hundred dollars, you don't know how it spends tokens. Same as in ChatGPT. You don't fully understand how tokens are spent. And who knows how they're spent? Here's, by the way, an interesting thing. A month ago ChatGPT sent me: "You've been gifted sixty thousand tokens." I asked ChatGPT: "And roughly how much money is that?" They wrote: "Well, a couple of thousand dollars, most likely." Then I have a question: how much will my tokens cost, which it shows in total as sixty-four billion tokens, right?
With six zeros added. Very strange. And so, well, actually, let's put it this way, to cover everything: there aren't tokens there, there are credits, and credits correspond to a certain number of tokens, yes, so it's calculated a bit differently. But the point is that they gifted these sixty or however many thousand credits. I calculated, and I was told that's two thousand dollars. And then the next week I get an email saying they distributed credits across the different plans incorrectly.
And I had just switched to ChatGPT Pro for five hundred dollars on the day the plan was released at the conference. And they write that they distributed credits incorrectly on Pro plans, so they want to compensate. Here are twelve or however many thousand credits for you, and that corresponds to roughly a thousand dollars, or eleven hundred dollars. And I'm like: wait, the chat told me sixty thousand credits is two thousand dollars, and you're sending ten or twelve thousand, and that's a thousand dollars.
Essentially, nobody understands what this really is and how it's then spent inside. What's more, you also don't know how it's spent on your task. Sometimes you can launch some task, for example, and your task gets solved very easily, quickly, right away, with super quality. And sometimes you can launch a certain task, and you don't understand, you absolutely don't understand why this task isn't getting solved on the fifteenth try, when you're already using the coolest and most expensive models.
What is the fundamental difference between ChatGPT Dots and all three systems mentioned? Well, actually, I would still compare it with you, for now, with Claude Duet.
Now I'll tell you another fundamental feature in terms of choosing systems. Claude Duet eats your tokens, and if your tokens have run out in Claude overall, then you've stopped, and you're dead. I think this is a very serious oversight by Anthropic. It shows that Anthropic isn't focused on making their artificial intelligence model an everyday assistant for an ordinary person. That's why even when they talk about Claude Duet, they say that Claude Duet helps sort out customer invoices.
Like, working through QuickBooks or through Stripe, that it helps work with customers in general, use the history of contacts and deals, for example, with an integration through HubSpot, yes, or through Salesforce, or signing documents, through DocuSign, for example. Or they say that it lets you, with marketing, for example, prepare campaigns for you, in Canva. That is, it's that kind of orientation. They keep showing cases and work through third-party systems. They say: "You can work with your online store through Shopify." And what does working mean?
It's managing, in essence, your business in terms of selling goods, or doing some serious analytics in terms of markets and so on. ChatGPT Dots, at its base, in general, is about the work of an ordinary person. And for an ordinary person, just like your ordinary messenger, it can't run out of limits. It can't be that you're using Telegram or WhatsApp, and at some point either iMessage or SMS tells you: "That's it, you've run out of limits." With SMS it used to happen, yes, when SMS were expensive. I don't know, I connected my first phone on this type of network in two thousand one, right?
Back then, twenty-five years ago, that was a thing. I counted my SMS. Before that there was a big handset on a different standard. I think you couldn't even write SMS there, if I'm not mistaken. You probably couldn't. It was a different standard, right? And, well, that was around the year two thousand, ninety-nine, ninety-eight. Well, before, once again, I already had GSM. I got a phone in two thousand one. And a Nokia 8310, I remember it like it was yesterday, was my first phone specifically on the GSM standard, yes.
And I've even forgotten what that old standard is called. Well, if you remember it, write it, yes. That handset, my big one, I remember coming to university with it, with that huge antenna. The battery didn't last long. And it usually had a landline number on it. Whereas now, of course, you can't have a situation where your messenger is limited, yes, or your email is limited. Well, that's fundamentally an abnormal scenario. In simple use, they can limit you for too much email, they can limit you for too many files, and they can limit that.
Well, even too much correspondence in Telegram or WhatsApp, they generally don't limit you. There are some limits there, by the way. How many messages you can send at once, the same message to different users, for example. Telegram has a limit, but basically a person doesn't even have such cases, if he's with different, and with different people, I don't know whether I can write to thousands of my friends at once, but probably an ordinary person doesn't do that anyway, right?
And you're unlikely to do that in Dots or in Duet. How would you even do it? You have a single input context window, and you'd need to enter a bunch of things and launch a bunch of things in parallel somewhere. And notice, precisely, that Dots and Duet will start creating sessions, and, and Dots will get limits too. Because if Dots uses not the chat, but uses Work, a product that's still very strange to me as a concept, or uses Codex, then your tokens will be eaten. So, in Claude Duet, it uses tokens. Yesterday I sent it, well, two or three messages. I say: "How are you spending tokens?" It says: "Yes, I've spent roughly $2."
I say: "And how could you have spent $2 if we're just using my weekly limit right now?" It told me: "I roughly recalculated how many tokens I spent." And I'd sent, you know, a series of messages there. "Look at my task, how often it runs and how much money you've spent." It was roughly like that. I think I sent two or three screens of messages there. Each one tells me $2. You're like: "How $2? How is it possible that it spent that much money at all?" And now the most important thing.
What is the biggest fundamental difference between systems such as GrokBot, Meta Muse, ChatGPT Dots or Claude Duet? The biggest difference. We have to understand that the appearance of these systems already marks that it's not a question of choosing a model, it's not a question of choosing how, at what level, the model works. Whether it works in light mode, or in medium mode, or in maximum mode. Remember, at some point we had reasoning settings, right? How it thinks. Or whether it thinks in pro mode, or in maxi, and so on. You can still get confused by this.
It's not a question of even controlling the speed, although you can, by the way, launch through Dots, say, on the five-hundred-dollar plan you can launch 8x speed to solve a certain task. I noticed that I stopped thinking about models while working through Dots, and yesterday I was doing a fairly non-trivial task, for a 3D printer. I need to print something there for one of the basketball hoops. I have two hoops there, and on one hoop it's impossible to buy, the latches, one set.
By the way, maybe I'll tell you about this case later. Max Grigoriev said yesterday that it would be a cool case to tell on the channel. We were walking together just then, discussing how it would be printed on a 3D printer. So, I noticed that I no longer choose a model. That is, I use ChatGPT Pro less and less. By the way, I think for OpenAI I'm definitely a loss-making person. My account always was, too, and I started using pro mode less, because Dots appeared. Because I'm a fan of always asking every question in pro mode, and that's an expensive mode, well, really a complex mode. Of course, I keep using it, but overall.
You have to understand that these systems are an entry point into an infrastructure, and comparing micro-things becomes meaningless. At some point you could compare computers with each other in terms of RAM, or in terms of processor speed. Then they started comparing screen quality, started comparing the weight of these computers. But if I buy an Apple phone, I just buy it because a new Apple phone came out, not because I simply compared the specs. If all people compared Apple on specs, Apple wouldn't be the leading phone-selling company in America. Or I buy an Apple computer, well, because I'm an Apple person.
That's the story, because, what am I going to compare? Well, there are things that are better, there are worse ones, there's something else somewhere. And when someone starts telling me that Huawei or Samsung has some cooler technology, well, I go, okay, cooler, but overall, as a whole, I'll choose the Apple infrastructure. And such systems as Dots and Duet, I really hope they will make these systems into what I've just been describing about them. And because I'm telling you, on the one hand, a piece of what exists now, and on the other hand, I'm giving you some cases that, which this system basically isn't even designed for, right?
What it means for the system to become an operating system. That is, in essence, you open your phone, and there's only Dots and basically nothing else. It's your central entry point. And then you can open inside Dots a messenger, well, like a ChatGPT messenger, or launch a separate session for something, or separately open some design app. But it's all within, for example, OpenAI's infrastructure, or all within Anthropic's separate infrastructure. And, of course, it's not just an operating system, yes, because you can't call it that.
We'll see, by the way, a new name, a new word. A new word will appear for what this will be called. In the past, with a computer, they said very clearly: here, there's the BIOS separately, yes, there's an editor here, standing separately, there's an operating system, standing separately. Back then this setup was strong, like, I don't know, in '93, in '94, '95. That is, whoever bought himself a computer and dealt with installing Windows or Linux knew this very clearly. Knew all these individual quirks inside.
Some new term will appear, definitely. And here I even feel close to Elon Musk's position, as he told my friend, that it's not about there being a chat. What there will be is the question of what it will be called, what kind of system it will be. And that's why when you choose Claude Dot, Claude Duet or ChatGPT Dots, when you ask a question, or ask questions: "And can I use Hermes, since it's open source, and connect something through it? Or I want to use DeepSeek agents, yes. Or Kimi Moonshot just made an update there, yes, I want to use that. Or I want to use..."
They did update, by the way, but notice, they, Moonshot itself wrote to me, even before we maybe talked about them, when they were releasing Kimi K3, and they said: "We're at the level of Fable and Astra." But that's just an absurd claim. Well, in some tests, maybe, but basically you're not Astra and not Fable, you're basically not OpenAI and not Anthropic. What's more, Anthropic and OpenAI obviously have models far cooler than Astra and Fable. Well, Fable is the Mythos family, well, cooler than Mythos.
Sam Altman recently says in an interview: "Of course, we have cooler models." But it's, it's obvious that they have cooler models, they have more access to these serious top-class models. And you, accordingly, choose for yourself and are already increasingly choosing a certain infrastructure. And there's a rule that I strongly put out there. On our site there's a list of some trends I talk about, or my thoughts, or rules for working with artificial intelligence. There are more than twenty of them, you can go and look, we'll give the link, yes. Choose one main system for yourself, use all these cases right inside it, and somewhere a second, additional one, if you have money for it.
If you don't have money for it, you'll always feel a certain micro-thing. But a third or fourth system is hard to use. And I have access to a paid plan. I have a paid plan with Google, in Gemini, Antigravity, everywhere. Well, I launched something there, looked, that's it. Well, not my system. That is, if you get used to a certain infrastructure, you get used to a certain infrastructure. And which infrastructure will you choose? It's a very interesting, interesting topic. Write here in the comments what your choice will be.
And once again I ask you, support the channel, don't forget to hit the bell, subscribe if you haven't, even if you watch through notifications. Give the channel a like. By the way, tell us what time you watch. It's interesting for us specifically for active viewers. Not all viewers watch at the same time. And from which countries. By the way, we haven't read in a long time which countries people watch our channel from. We get very different requests. For speaking engagements, I see, people come from different countries, well, people reach out. Don't forget that we also have a site, among other things. Write comments, share with your friends.
And I'll see you in our upcoming episodes. There'll be a very interesting episode in the next few days. Our joint weekly podcast. Come join us there, we have a big team there.