How much does it actually cost today to use powerful systems such as ChatGPT, Anthropic, or Gemini? What do they really cost? And if using a system like Codex can cost one hundred thousand dollars a day, who will have access to these systems in the future? That is what we are going to discuss today. Hello, everyone. This is ToTheMoon. What happened? I opened Codex and accidentally came across a new profile section showing how many tokens I had used. Maybe the section had already been there; I had simply never looked at it.
I have subscriptions to various systems, and there are also cases where I pay the full token cost myself. I have said many times that, in my personal view, my work inside Codex and ChatGPT is unprofitable for the company. I have had the $200 ChatGPT Pro subscription since it appeared several years ago. I was among its first users, and OpenAI even sent me a pen. The page says that I have been a ChatGPT Pro member since 2024. It is funny: a pen. There have been endless discussions about whether my usage is really unprofitable or whether requests are actually cheap.
Today I opened this new personalization section. We will not talk only about cost. I also wanted to touch on memory: the major changes in what the system remembers about you and in its ability to analyze all your chats at once, regardless of whether you saved them or whether you are working in one context window or several. I also wanted to discuss what is called AI dreaming, which is translated into Russian as “dreaming,” and whether that is even the right term. But the subject of price is enormously important to me, and I think it matters to everyone.
What happens next? Who will have access to these systems? Jensen Huang has said that if a programmer in his company earns half a million dollars a year but does not spend half a million dollars on tokens, that is a bad sign. At the same time, we see the situation at Meta: Zuckerberg told everyone to use as many tokens as possible and work heavily with the systems. After the company used several trillion tokens, it began imposing limits because the bill was reaching, or could reach, billions of dollars over some period.
Then I saw an interesting number in my own account: I had used forty billion tokens.
At first it was just a number. I have said many times that no normal person can think in tokens. It is strange to report a result in tokens at all. On OpenAI’s page there are “tokens,” “peak tokens,” “current streak,” and “longest streak”—labels that do not even make it clear what they mean. When I saw forty billion tokens, I initially thought the figure might include ChatGPT. It turned out that it did not; it referred only to Codex. That was especially strange because Codex said I had 3,270 chats, and I do not believe I created 3,270 chats.
It also said I had used 1,088 skills, although I do not personally spend much time creating skills. Many people now describe how they configure skills and set up all kinds of things. I do not like that approach. I advocate building AI-native software: the artificial intelligence does everything for you. You do not live in terminals; the system makes decisions, writes the code, and completes the work itself. I believe that is the only way to create something genuinely serious and make real progress.
I am not talking about professional programming, where a developer sits down to implement a specific module or a particular part of the code. When I saw forty billion tokens, I had a rough idea of the pricing, so I asked ChatGPT how much I would have paid if I had bought all of those tokens instead of receiving part of them through a subscription. I cannot say that I have personally spent tens of thousands of dollars on tokens, but I have spent amounts I did not pay much attention to.
Had the total exceeded ten thousand dollars, I certainly would have noticed. Many different systems charge me money, including API services for things such as audio recognition. I also pay token costs, but I set daily limits—fifty, one hundred, one hundred and fifty, or
two hundred dollars. Those are reasonable amounts. I asked about forty billion tokens and received an answer that made me uneasy, so I decided to share it immediately. The rough middle estimate was between three hundred and five hundred thousand dollars using something around GPT-5.5 High. I told the system that I use Extra High, and the upper estimate became about $1.2 million. I decided to investigate. Codex lets you view daily token usage, and I found several striking dates.
Over July 14, 15, and 16—three days the previous week—I used fifteen billion tokens. I remembered that I had been complaining about those exact days, when GPT-5.6 Sol became active. I was using it in Ultra mode and had launched it on many tasks. I also had a large number of accumulated ChatGPT resets, and ChatGPT itself periodically reset my limits. I think I mentioned in the previous podcast that the limits kept resetting. I spent a great deal of time using GPT-5.6 Sol, but I did not obtain a result.
Even before I saw the token figures, I had stated very clearly that I felt sorry for people who had bought GPT-5.6 Sol usage through the API. The internet can say endlessly that it is a brilliant model and produces wonderful answers. It does do some impressive things in chat, and I can even see some decent programming results. But at the moment, although the model will naturally improve and someone watching this later may tell me that GPT-5.6 now works well, I am making a broader point about the reality we are entering.
That point will not become outdated. Those fifteen to twenty billion tokens—fifteen billion over three days—have a theoretical maximum cost of roughly half a million dollars if we are talking about Sol, and I know with certainty that this usage was on the Sol model. A more typical estimate is still around one hundred and fifty to two hundred thousand dollars for three days. Once again, GPT-5.6 Sol did not produce a meaningful result for me.
Worse, it stopped a great deal of work that had already been in progress and disrupted many of my results. Anthropic’s Fable was different. You know that I am a major admirer of OpenAI and ChatGPT, and for now I will probably remain one because I still conduct my main conversations there. I have an enormous number of chats, systems, and analytical projects.
With GPT-5.6 Sol, larger context windows, and more reasoning capacity, the system can perform serious analyses, and there is much that I like. But here I agree with Elon Musk that Anthropic has begun capturing an extraordinary market. Which market? Anthropic is not winning global name recognition; a billion people do not use it. OpenAI will probably still be worth more than Anthropic at an IPO. Yet my own view is beginning to shift because of some remarkable things OpenAI is doing.
OpenAI remains an extraordinary leader because it has a billion users and owns the main consumer chat product. But in system development, Anthropic has become a very serious machine. If we are talking about leadership in creating AI-native software—software or even entire companies built by artificial intelligence—Anthropic may win that category. I recently made an episode explaining what AI-native means, including the partly philosophical implications. ChatGPT may correct course and produce a different result, because today it is extremely difficult to judge which system is genuinely better or worse.
Moonshot released Kimi K3, and people immediately wrote that everything else was obsolete because Kimi K3 was much cheaper and supposedly brilliant. I read those claims and wonder how anyone tested it, how they verified it, and what the system really is. Chinese companies price for their own market. They cannot charge the same token prices as Anthropic or OpenAI because nobody there would buy the product at those prices, except perhaps in markets where OpenAI and Anthropic impose strict sanctions or access restrictions.
Chinese companies charge what their market can bear. But Anthropic and OpenAI are loss-making companies. When one subscriber can burn through one or two hundred thousand dollars’ worth of tokens in a day under some plan and then have the allowance reset again, you have to ask what is happening. OpenAI recently ran a promotion on X: users could post how Codex helped them and receive one hundred dollars in credits. Ten thousand people were each offered one hundred dollars, so OpenAI was giving away one million dollars in tokens.
Against subscription-level token consumption, that is almost nothing. What can I launch for one hundred dollars? What can I actually complete? One of my earliest tests used GPT-5.5 Extra High. I bought one hundred dollars in credits, thought the budget might cover a broader task, and watched it disappear within minutes without producing a result. When a task costs one thousand dollars, you have to think carefully about how you write the context and describe every detail.
At that point it is no longer an AI-native system; it starts to resemble another form of programming. When I look at the spikes on July 14, 15, and 16—five billion tokens in a day—I understand that the tokens were not consumed only by me. They were consumed by an unqualified system. Who is going to pay for that? Who decides who can use it at full price and who receives subsidized access? What kind of people will qualify? Anthropic has now given every teacher in America free access to its Pro version.
It already had subsidies for teachers, as did OpenAI. You look at free access to a twenty- or thirty-dollar plan and ask whether that user is profitable or loss-making for the company, and what benefit the company expects in return. Current token costs suggest that Anthropic cannot subsidize usage to the same extent as OpenAI. I have not seen Anthropic permit the same volume of token spending. And unlike ChatGPT today, Claude chat is tied directly to the token allowance. When you sit in Claude chat, you consume the same tokens.
I think OpenAI will eventually do something similar.
It will be difficult, but once ChatGPT and Codex share a single token pool, access will suddenly shrink for everyone who uses the system intensively, and serious usage will become expensive. I understand that my own work could cost hundreds of thousands of dollars a month. Am I prepared to pay hundreds of thousands of dollars for my current use of ChatGPT and Codex? Honestly, no. I am prepared to pay something—perhaps up to ten thousand dollars—but probably not hundreds of thousands.
And how willing are you, or am I, to change the way we use Codex, Claude, or ChatGPT after becoming accustomed to high quality? Anthropic released Fable and initially said it would be temporary: look at this excellent model. As of yesterday, it has made Fable permanently available on the stronger, more expensive plans. There is a fifty-percent limit, but the model is available to me on the Max plan. Why? Because once you become used to auditing your system with Fable, you begin to depend on that capability.
I have not yet become accustomed to Sol inside Codex, although I am probably becoming accustomed to Sol inside ChatGPT. How do you see this? What do you expect to happen to access and pricing? Everyone should think seriously about it—not merely in passing, but as something that must be understood—because people underestimate what is happening and do not know what anything actually costs. Someone may say OpenAI exploits customers who buy tokens, but look at the company’s expenses.
These companies bear colossal costs. Consider why Google restricted the amount of Gemini capacity it would sell to Meta: Google reportedly said that Meta’s demand was consuming too much capacity and that the resources could instead be allocated to other customers or uses. Will high-quality artificial intelligence—genuinely capable, intelligent systems—really appear everywhere, in every assistant, car, and robot? What will that cost, and who will pay? Others point to China and to research claiming that companies can build highly capable systems that use few tokens and cost users very little.
Will that really hold? Then why is OpenAI deploying a model such as GPT-5.6 Sol, and why is Anthropic deploying the expensive Fable while still giving people a high-quality product? I can accept the argument for Fable, although I still have major questions about price, quality, and final effectiveness. I had wanted to connect this subject to memory, but decided not to blur the discussion. I will record a separate episode about memory and AI dreaming, to be released in a few days.
I have already made a strong episode about memory that you can watch in preparation. We will also introduce new authors on the channel. Why? Some viewers want more content and a wider range of perspectives. They do not want content only from me, Ilnar, or Tanya; they want broader and more varied approaches to artificial intelligence from different kinds of people. Anna Volchek, my sister, has already appeared. She is deeply involved in AI. Many viewers ask what a person does and who they are before deciding whether to watch them.
But what do you really know about me? Most people who watch this channel do not know me; they know only fragments. Some viewers wrote that every new author must be introduced. How should a person be introduced—in one sentence? We live in a world where people increasingly want to know only a single sentence about someone. I hope you will enjoy the episodes and that our regular audience will be able to choose the authors they like. Another question is whether companies hiring people who supposedly “understand AI”—prompt engineers, people who can build in Codex, configure agents, or whatever new title appears next, perhaps “AI-native systems specialist”—can evaluate them by cost.
How would a company measure how much money each person consumed and why? The system is opaque, inaccessible, and extremely difficult to understand. What happens next? How will any of this work? It is astonishingly unclear. Share your thoughts in the comments, and we will meet again in the next ToTheMoon episodes.