There was exactly an hour and a half left before OpenEA was to launch the much-hyped GPT6 Sol and GPT6 Luna. Then , 90 minutes before—just 90 minutes, folks—Antropic quietly launched Cludopus 5. 5.
Looking at it from the outside, you might think it was a marketing ploy, an attempt to steal clicks from the competition and get noticed before all the attention went to OpenAI. But looking at what Antropic launched, at the model's price, and especially at how it was designed, there's a much more interesting interpretation. The competition probably wasn't about hype, but about the companies 'infrastructure.
Because while everyone is still obsessed with discovering which AI responds best, which reasons best, and which earns the most benefits, another war is starting behind the scenes: the war over how much it costs to put an artificial intelligence to work on its own. And to understand why this can directly impact your routine, your job, and your wallet , we need to dismantle the official propaganda a little and look at the numbers. If you open Antropic's page, you'll be faced with the usual corporate spiel.
A pretty table with benchmarks, comparisons to other models , phrases saying that the Opus 5. . .
Intellectual, complex, programming, agents—all sales talk, right? Because saying the model is revolutionary for work doesn't say much about how it performs in the real world. But amidst these numbers, there's a much more interesting detail: the price.
The Cloud OPO 5. 5 costs $ 4 per million tokens to enter and $ 2 per million tokens to exit. The previous OP 5 cost $ on entry and $ 25 on exit.
In other words, the price of tokens has dropped by 20%. But Antropic says that in practice, the typical cost to perform certain tasks can drop by nearly 40%. Because this new model also achieves the result using less processing power.
And there's an even more interesting number: the memory read cost in cache, which has dropped from 50 cents to 20 cents per million tokens. That's a 60% drop, right? And if you have no idea what tokens, cash, or APIs are, don't worry, because the important thing isn't memorizing these numbers.
The important thing is understanding why a company like Antropic is fighting to make these numbers fall. This price cut isn't charity or a birthday promotion; it's the mathematics of autonomous agents. Until now, most people have used AI as if it were a talking encyclopedia.
You go there, type a question, and the system answers. Then you ask it to correct a text, it corrects it and stops. Then you ask for an idea, it gives you an idea and waits .
It's a purely reactive relationship. You speak, the machine answers, you speak again, the machine answers again. But when Antropic emphasizes deentic coding and autonomous work so much, the message is different.
The idea isn't just for you to communicate better with AI, it's for AI to be able to continue working even after you've stopped talking. So, instead of asking, "Help me write the code for a shopping cart? " , you say, "I'll set up an online store for you, create the database, write the cart code, connect the payment system, test everything, get it all up and running in the cloud, and then you're gone.
" To make something like that work , the AI needs to work in a loop. It writes a piece of code, tests it, an error appears, it reads the error message itself, looks for a solution, changes the code, tests again. If another error occurs, it repeats.
And this cycle can happen dozens or hundreds of times before the project is finished. And that's where those cents that seemed completely irrelevant start to become huge, right? If each cycle is expensive, leaving an agent trying to solve a problem for 2 hours can become prohibitive, right?
Now, multiply that by hundreds of agents, then by thousands of employees, then by an entire company. That's when you start to understand why price has become a strategy. Because an agent that works for a long time needs to keep consulting context, files, instructions , information that it has already seen before.
So, the cheaper it is to keep this context available, the cheaper it is to have the agent working for long periods. And that's why this price reduction matters so much, not for the person who opens the cloud once a day to ask something, but for the company that wants to execute thousands or millions of automated tasks. And pay attention to this part, because this shift from chats to agents is one of the things I'm following most closely on this channel.
So, if you like this type of content, subscribe, activate the bell so you don't miss the next videos, and if you want to support the channel even more, consider becoming a member, okay? Now back to AP 5. 5, because a low price alone doesn't solve the problem.
And there's something that no pretty little benchmark can completely erase. The more autonomy you give an agent, the greater the damage when they make a mistake. Then a problem appears that I like to call a cascading error.
Because the idea is like this: Imagine you deliver a project with 20 steps to an agent. Then , in step number three, they misunderstand something. But nobody's there monitoring what he's doing.
So he continues and then he does step four on top of that mistake. Then he does C, 10 , 15. When you come back, he's finished everything.
The system looks nice, everything seems ready, right? But the whole structure was built on top of a wrong decision that happened way back. It's like building a building and discovering on the 10th floor that the foundation, the base, was crooked.
And that's one of the big problems with autonomous agents, right? It's not just AI making mistakes. AI already makes mistakes today.
The problem is having enough autonomy to continue working on its own mistakes. Because if the system freezes and tells you, "Look, I don't know how to do it," great, you know where the problem is, you know it won't be able to do that. What gets complicated is when it keeps working, keeps building, and delivers something that looks perfect.
It's just the appearance of perfection, right? Because when you try to use it, you'll see that it wasn't perfect at all. And that's precisely where that somewhat abstract conversation about security starts to become much more concrete.
Antropic made quite a stir about the security improvements in OP 5. 5. According to Antropic, this was the model that performed best so far in their internal behavioral evaluations.
And they also say that it became less prone to performing actions that are difficult to reverse or exceeding the limits defined by the user. And look how these two things are connected, right? Because the more Antropic tries to transform the cloud into an agent capable of working on its own, the more it needs to prove that this agent knows how to respect limits.
Imagine an autonomous agent operating within the server of a large company. It can gain access to databases, paid services, internal files, infrastructure, APIs, and various cloud-based resources. Now imagine hypothetically that this agent enters a loop because of an error and starts consuming resource after resource for an entire weekend.
On Monday, someone goes there, opens the dashboard and finds a bill with tens of thousands of dollars. Who will pay that bill? Or if it deletes something it shouldn't, accesses the wrong service, makes an unauthorized change that nobody requested, or takes a decision based on information that it invented itself.
And that's when the conversation about a possible machine rebellion starts to get closer, right? It becomes a much more mundane discussion, right? Who gave permission for this?
How much can this agent spend ? What can it change? What can't it touch?
When does it need to stop, and at what point is it obligated to call a human being? Because creating a virtual intern who never gets tired seems wonderful, right? Until that intern gains access to the company's credit card and decides to work alone on Saturdays and Sundays.
But now we come to the part that is of real interest to most people. What does all this change in the life of someone who isn't a programmer, doesn't understand servers, and just wants to guarantee their monthly income? The concept of office work begins to change.
In recent years, we've become accustomed to seeing AI as a friendly tool, right? Like an intelligent spell checker, right? An assistant to summarize a huge email, translate something.
That was the era of the digital assistant. You did the work and called on AI when you needed help. With agents, the logic starts to change.
You deliver an entire task and then monitor the result. And this greatly changes the type of skill that becomes more valuable. If before you needed to know how to assemble that entire spreadsheet, perhaps now you need to know how to explain exactly what you want to configure the agent for, check the result, and notice when that agent makes a mistake, right?
If before you wrote each part of a code, now you may end up having to coordinate several agents that write, test, and correct parts of that code. Does this mean that all office work will disappear tomorrow? No, right?
And the problem of cascading errors itself shows exactly why. Because the more responsibility you give to the machine, the more expensive an error that nobody noticed can become. But it's also difficult to ignore this, because operational tasks that previously required human hours can now be performed in increasingly cheaper automated cycles.
And if your main role within a company is to perform repetitive tasks, especially junior-level tasks, this change deserves a lot of attention, okay? Because you don't need to compete with perfect AI. The machine doesn't need to do everything better than you to shake up the market.
It just needs to perform a sufficient number of tasks well enough, fast enough, and cheap enough. And that's where the difference between those who know how to manage these systems and those who try to compete directly with their speed begins to grow. Now let's go back to the beginning.
90 minutes Cloud OPO 5. 5 and then GPT6 Sol and GPT6 Luna. Because there's something very interesting when you put these releases side by side.
Antrop wasn't the only one that came in with aggressive pricing; OpenAI did too . And then that story, it's no longer just Antrop that lowered the price of Claude, because something much bigger is happening. The A-list companies are starting to compete not only on who has the smartest model, they're competing on intelligence per dollar.
How much work can this model do for the lowest possible cost? Because if AI is only going to answer your question occasionally, a few dollars' difference per million tokens doesn't seem like a big deal, right? Now, if it's going to work all day, if it's going to make thousands of calls, if it's going to read documents, use tools, test code, fix errors, talk to other systems, repeat all of that thousands or even millions of times, every penny matters.
And that's where price becomes strategy , reliability, security—it all becomes strategy, because we're seeing that the next phase of this war won't be won simply by AI, which puts on the most impressive demonstration for people to see. There may be a much bigger dispute happening behind the scenes of companies 'systems: on servers, in APIs, in internal tools, in processes that happen thousands of times a day without anyone noticing. So now there's a war to decide who will provide the engines that make an increasingly large part of the digital economy work.
And that's why those 90 minutes of difference are so interesting. I can't say for sure that Antropic chose that specific time to block OpenAI. But looking at the product they've now launched, it's easy to understand why launching earlier made so much sense.
They managed to get OP 5. 5 into the conversation before all the attention turned to Sol and Luna, and most importantly, they managed to send a very clear message to the competitive market: "We're here too, our model has become cheaper, it was designed for agents, it's ready to integrate into your infrastructure. " However, OpenAI's response shows that this isn't a strategy exclusive to Antropic, it's not just them doing it, it's just the new phase of the race.
Because now what really matters is how much it costs to let this intelligence work alone. And this change seems small until you realize what it means for human work. Because if the cost of delegating entire tasks to agents continues to fall, the human value increasingly shifts from pure execution to the ability to decide what needs to be done, give good instructions, coordinate tools, and take on decisions that the machine shouldn't make alone.
So, the bar for humans has been raised another notch, not because these agents have become perfect, right? They haven't stayed yet, at least. The truth is, an ever-increasing amount of work can be delegated to systems that work in an infinite loop, cost less and less , and can retry as many times as needed .
So, the most important skill in this next phase isn't competing with the machine in execution, because you'll lose, it's learning to be the conductor of this automated orchestra, right? Because companies are already fighting to decide which of their agents will work alone behind the scenes. And now I want to know what you think.
If you think these agents will primarily increase the productivity of those who work, comment "ally. " If you think they will replace a huge part of office work , comment "replacement. " I want to know what you think about this.
And until the next leap.
