OpenAI puxou o tapete e todo mundo vai fazer o mesmo

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: OpenAI puxou o tapete e todo mundo vai fazer o mesmo · Channel: AI ProgBr · 13:38 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, AI models, Local and open-source models

Summary

The video discusses OpenAI Dev Day, focusing on Dots, OpenAI’s new agent-like product, and the launch of a $500 Pro plan. The presenter argues that reducing the limits of existing plans while introducing a much more expensive tier could signal a broader industry trend toward higher AI costs. He questions whether improved model efficiency will truly offset reduced quotas and suggests that users explore alternatives, including smaller and local models. The video also promotes learning how to get more from lower-cost models and closes with channel updates.

Key points

OpenAI introduced Dots, described as a product similar to Hermes or OpenClaw that runs OpenAI models; its usage is initially covered by OpenAI rather than the user’s quota.
Dots access is described as available to Pro, Business Premium, and Enterprise users, with some possible limitations in the European Union.
The presenter highlights the new $500 plan and says the existing $200 plan’s usage allowance is set to decrease for new subscriptions.
He argues that the new plan structure offers little added value compared with the previous limits, while making high usage more expensive.
OpenAI cites more efficient and less expensive models as a reason for changes, but the presenter questions whether this will compensate users for reduced quotas.
The presenter expects other companies may follow similar pricing changes, based on past industry trends.
He recommends considering alternative models and learning to use smaller models effectively, rather than relying exclusively on the most powerful offerings.
The video notes advances in local model performance and mentions examples of running models on the presenter’s hardware.
The presenter promotes a course about harnesses and getting more from cheaper models, and previews upcoming channel content.

Chapters

Start	Chapter	Summary
00:00	OpenAI Dev Day and the launch of Dots	The presenter introduces the event and describes Dots as OpenAI’s agent-like product, including its access and quota arrangements.
02:02	New plans and changing usage limits	The video examines the $500 plan, compares its limits with existing tiers, and explains the announced reduction to the $200 plan’s allowance.
06:14	Pricing rationale and questions about value	The presenter discusses plan availability, the Ultra Fast model tier, and whether cheaper, more efficient models make reduced quotas fair.
09:25	Alternatives to relying on the largest models	The presenter predicts that AI services may become more expensive and encourages viewers to consider other models and improve their use of smaller ones.
11:01	Harnesses, local models, and closing updates	The video promotes strategies for working with smaller models, shares local model performance examples, and closes with a course invitation and channel news.

Quotes

Dots will be like a Hermes, an OpenClaw, but from OpenAI — 00:30

R $ 500 is 10%of that for you to have access to AI. — 03:05

lowering the limit of a plan and creating a plan that’s more than double the price — 05:12

Probably all the other companies will follow. — 09:57

The way to avoid depending on these large, excellent models, which are becoming unviable, is to better understand how you can get more out of smaller models. — 11:01

Concepts and entities

Dots, OpenAI, GPT-6 Astra, Hermes, OpenClaw, Qwen3.8, DeepSeek, Opus, Anthropic, Sonnet

Full transcript (original language)

00:00 Okay, folks, today was OpenAI Dev Day, and we have some news in the world of artificial intelligence. I wish it were good news, but it’s not that great. But anyway, let’s start with the positive side. Let’s go to the screen so I can show you. The OpenAI Dev Day. It’s an event, right, from OpenAI where they talk about new products, new releases, and new changes. And an event like this has two motivations. The first is to get closer to what the competitors are doing and also to get ahead of the

00:30 competitors, dictating what the industry will probably do from now on. So let’s start slowly with the good part. OpenAI launched Dots. Dots is OpenAI’s MUS. I even wanted to make a video about Music and haven’t had the opportunity yet, but basically Dots will be like a Hermes, an OpenClaw, but from OpenAI, running OpenAI’s models. It will even run GPT6 Astra. And when I heard that, I got a little scared, because the GPT6 Astra is an excellent

01:01 and expensive model, very expensive. So I said, “Okay, wait a minute, if he’s going to use that model, I’m not going to use that, because it’s going to consume all my quota.” And then, to my surprise and relief, the quota used by the Dots on the GPT Astra won’t come out of your account, it will come out of OpenAI’s account. Moreover, for this month, I believe it is somewhat unrestricted, without limits, so they can see how that consumption will be. And there’s something funny, because I saw a post from the Meta leader talking

01:32 about Music, which is similar to Dots, and he compared several icons from the companies of Achatopic models, etc. And they all looked, how can I say, like a mess, let’s say, right? Except for Music, which is Meta’s. So, I read this this week, and when Open AI launched Dots today with this icon, I couldn’t help but remember it. So let’s start with who will have access to Dots right away. At this moment, it will be available to those who have the Open AI

02:02 Pro plan. So, until yesterday we had the GO plan, which I think has ads and everything, the Plus plan, which is the $ 20 plan, and the Pro plan, which is $ 100 and $ 2. Calm down, I don’t want to give spoilers here. But well, that’s what we had until yesterday. So, the Pro plan has access to Dots at this moment. I don’t even know if everyone already has it, but the Pro plan, the Business Premium plan, and the Enterprise plan will have access to Dots. There are, I think, some limitations for the European Union, but in general, this is the scheme. But well, Dots will run on your own

02:33 computer, on a private network, on a virtual machine, in the GPT chat cloud. If you’ve used Hermes, Opencore, or Music before, it’s more or less the same idea. Now let’s get to the point. As I was saying, until yesterday we had a Go plan for 8, a Plus plan for 100, and a Pro plan for 200 until yesterday. Because today we have a plan for 500. Wow, if I already thought that was a bit steep, 500 is absurd. Imagine for a Brazilian, right? 500 is what? R $

03:05 2,500. Let’s say the guy earns a good amount of money. R $ 25,000. R $ 1,000 a month would be, wow, an excellent salary in Brazil. R $ 500 is 10%of that for you to have access to AI. If you earn R $ 10,000, we’re talking about 25 %of your salary. It’s very expensive. And if you have the $ 200 plan, you might be thinking, “Oh, but I don’t need that. My plan is more than enough, “ but that’s where you’re wrong, that’s where the problem starts. What is the $ 00 plan? Well, it works more or less like this. The plans evolve in multiples. The base is the Plus plan.

03:38 The Pro plan, for example, has five times more limits than the Plus plan. Until yesterday, the $ 2 plan had 20 times more limits than the Plus plan. This is important. Why? Because the $ 500 plan has 25 times the limit of the Plus plan. And you might be thinking, that doesn’t make any sense. Why would someone subscribe to a $ 500 plan with 25 times the size of the Plus plan? If they can subscribe to the $ 2 plan,

04:08 which has 20 times the limit of the plan, almost the same thing, right? 20, 25, 200, 500. So, it makes no sense. Nobody in their right mind would sign up for a $ 00 plan to get 25%more. That’s right, nobody would do that. So what did they do? Your $ 2 plan remains 20 times the size of the Plus plan until next month. From next month onwards, your $ 2 plan will be worth 10 times the value of the Plus plan, that is, one to one. If you had 10 plans of

04:42 20, you would also have 10 times the limit of the Plus plan. Basically, now you only have the convenience of paying one bill, nothing more than that. There’s no longer a big advantage. And remember how I opened this video saying that this event has two objectives, right? To make OpenAI able to catch up with the competitors. In the case of dots…So Twitter, for example, has Grockbot, Meta has Music, they all ride the wave of success, like Hermes and OpenClaw. Now OpenAI has joined in and has its own Dots, but it is also here

05:12 to dictate exactly what the market is going to do next. And it’s dictating it in this way: lowering the limit of a plan and creating a plan that’s more than double the price for basically the same limit you had before. So, if you had a $ 200 plan, now the plan that will suit you is $ 0. Wait, I’ll give you another second to absorb that. 500. And I know how the internet is, okay? Right now someone here in the comments is already writing. He’s going to

05:42 delete it in 20 seconds, but he’s writing: “Igor, you’re talking nonsense . I see it on the screen. That’s what it says. There’s no $ 500 plan. So here it is. Right now, I have an account with a $ plan. So if I come to the upgrade screen, I can choose to upgrade to $ 200 or $ 500. Man, I never imagined that. Imagine paying $ 500 a month to have access to a tool that, if you don’t use it, it doesn’t matter, it’s over. Imagine you subscribe to

06:14 this, you’re going to travel for a week . That week cost you more than $ torn up, so, of course you can use your Dots , right, to leave it working there while you travel. Maybe, maybe that’s an option. I leave my Hermes, which is running on a Zima Blade, which is a computer this size, and using my local model here on my Mac Studio. Okay, let’s go. I I’m here looking at their documentation in Portuguese, and it’s outdated because they paused the $ 200 plan on September 10th, right? And now they’re back. If we switch the language

06:46 to English, we can see the documentation is already updated with the new 500 GPT Pro chat. A reason for this difference could be that this plan might not be available in Brazil yet. But what else is new in this plan? The Astra Ultra Fest, which will only be available in this tier. Only the $ 00 plan will have access to Ultra Fest. Ultra Fest, if I’m not mistaken, is Astra running at 300 tokens per second. That’s really cool, but it should also consume everything quickly, right? Look , this is important. You can re-subscribe to the $ 200 plan now. But

07:21 new subscriptions will have a smaller usage quota than before, okay? Supposedly, what’s their excuse? Here? Their excuse is this. I’m here on the API pricing screen, and if we look at GPT 5.6 Sol, it had a price of 4 per million input tokens and 20 per million output tokens. GPT 6.1 has half the input cost and half the output cost. Even caching the input has become much cheaper. So that’s their excuse: that

07:52 now the models are cheaper, right? They’re more efficient. So supposedly you won’t see a decrease in service. But in that logic, technically you will see a decrease or you will have a loss of value. Why? If before a plan of $, for example, served you well and met your needs within the limits, and the $ plan is five times the $ plan, right? You were there, okay, I pay 100, but I use 100, right? Which is fair. And I have much more limit than the person who pays 20. Now, if we consider that

08:23 with the same value you The efficiency will be doubled; the person who pays 20 will benefit because now they can consume twice as much. You, who pay 100 and were at the limit, are no longer at the limit; you continue paying 100. So, ideally, you would switch to a 50 plan. There is no 50 plan, so you will continue paying 100 because the 20 plan in the most recent GPT would be worth, let’s say, 40 in the older GPT. But 40 in the older GPT didn’t meet your needs . 100 in the older GPT is what met your needs. In other words, now you will pay

08:53 $, there will be leftover credit, there’s nothing you can do, you will lose that credit. If you continue with the same consumption pattern, you gained nothing, you only lost. The person who paid 200 and had 20 times the credit now has 10 times the credit, right? So they also gained absolutely nothing. This assumes that the models are twice as efficient. Efficient. So for this guy, maybe it’ll even out, let’s say. So these are the scenarios of the plans that already existed. But the fact that there’s a $ 00 plan makes me seriously doubt that this will stay that way, okay? It might last for a

09:25 week, two, maybe a month, but then we’ll see that it’s not quite like that , that the guy who was paying 200, even with more efficient models, will start to feel the need to migrate to the 500 one. Here, look, whoever already had the plan will continue with the 200 plan as it was until October 29, 2026. I’m recording this video on September 29, 2026, that is, exactly one month. After that, you’ll change to the smaller quota anyway. So, the fact that you, oh, look, this month here you

09:57 continue with your same quota, is another sign that this idea that the models are more efficient isn’t very real. Yes, and it’s becoming…More expensive. That’s just the first step. Probably all the other companies will follow. It’s happened in the past, okay ? And now with larger models, despite the logic that we have more efficient models, we have very good and efficient models, including local ones. I made a video here about the Quen 3.8 with 27 billion parameters, and the quality of that model for its size is absurd. And

10:29 I think that model is already enough for many people to use, including for code. But we also have other interesting models, Deepseak, V 4.0, V 4.1, a Quen 3.8 Next Flash, and Quen 4 is coming soon. So these are very interesting models. If you are dependent on large models, if you always depend on using Opos, always on using Astra or the best GPT available, maybe it’s time to start looking around and see other models as alternatives. And other models are also becoming expensive. The Kimica 3 is already much

11:01 more expensive than the Kimica 2.7. It’s better, it It’s bigger, but it’s also more expensive. And the way to avoid depending on these large, excellent models, which are becoming unviable, is to better understand how you can get more out of smaller models. This involves understanding harnesses very well, understanding how everything works, and creating strategies so that model knowledge doesn’t limit you. And these are exactly the topics we teach in our” E for Deves “course. This course was created precisely for this

11:33 moment, to help people prepare for a time when it becomes impossible to have a plan that can cover the entire month. If you have a $ 20 plan from Antropic, for example, you can’t develop with it. It runs out very quickly, unless you switch to SONET. But if you’re abandoning Antropic to use SONET, even though SONET is tough now, right? But let’s say before the last one came out, if you were using SONET, you could use any other open model, any other much cheaper model. When I say open model here, I’m not talking about running it locally. No, because that’s a whole

12:04 other story. Everyone thought I was crazy when I bought my Mac, my 5090. I’m almost able to run everything using only the local model, because the evolution of AI is coming from all sides, including the end users who run the local models. I remember running Quent.827 B on my 5090 at around 60 tokens per second, 70 tokens per second when it came out. Now, thanks to the MTP and so on, we are finally able to run it at a speed of 120 tokens per second. On my Mac M3 Ultra, I ran Quent

12:35 3.8 Flash Next when it came out at 30 tokens per second. Today I ran it here with a peak of 140 tokens per second while writing code. At 140 tokens per second. A model as good as Quent 3.8 Flash Next. You can use it locally, but as I said, that is another topic. For another day. If you want to see the quality of the tiny model like the Quent.827 B, 827B, take a look at this video here. And in this video, I’m running it on a 5090. But if you want to see how it runs on a much cheaper

13:06 motherboard, a Mark B70, also 32 GB, costing only 1/4 of the price, then take a look at this video here. If you want to learn more about extracting the maximum from harness and cheaper models , take a look at my course, okay? It has all the information for you to understand the proposal. And if it’s interesting for you, sign up there and I’ll see you inside. And of course, don’t forget to leave a comment below and subscribe to this channel, because there’s a lot of interesting stuff coming. Finally, our video with the two ARC B70s will be released. Soon I had

13:38 to assemble, look, everything is here for me to assemble, so I still have to assemble it. I had to assemble a new computer to be able to put this setup online and bring it to you. Here’s everything for you. So, thank you very much to everyone who supports the channel, who watches, who likes, who leaves comments, because it allows us to bring more interesting things here. That’s it, I’ll be signing off now, until the next video. Thanks!