Gemini 4 Argon na prática: por que ainda prefiro o Claude

Author: Rafael Quintanilha – QuantBrasil · Watch on YouTube

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: Gemini 4 Argon na prática: por que ainda prefiro o Claude · Channel: Rafael Quintanilha – QuantBrasil · 32:35 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, AI evaluation and benchmarks, AI models

Summary

The video reviews Gemini 4 Argon’s availability, pricing, benchmark performance, and use through Google’s Antigravity CLI. The host finds that Argon is competitive with Sonnet-level models, but reports slow performance, frustrating permission settings, and high usage consumption during a practical coding task. In his view, Google AI Pro’s limits are too low for professional daily use, while Ultra is expensive compared with Claude’s plans. He suggests Argon may still be useful for occasional tasks, such as making text sound more natural, and recommends self-verification to improve agent workflows.

Key points

Gemini 4 Argon is available to Google AI Pro and Ultra subscribers, as well as through Google’s developer offerings; free and AI Plus users do not get the same access.
The host says the model’s introductory API pricing puts it in a similar general range to some competing models, though total task cost can vary with token use.
Benchmark results place Argon in the competitive mix, around the level of Sonnet 5.5, but generally below Opus 5.5.
The host praises Google models’ natural-sounding writing and considers text review a possible use for Argon.
In the Antigravity CLI test, the default permission behavior repeatedly interrupts work; the host recommends enabling auto mode, while criticizing its reliability.
A 404-page coding task took longer than the host wanted, though he acknowledges that his unclear instructions and two codebases contributed to the delay.
The task consumed about 10% of his weekly limit, leading him to conclude that the Pro plan is not sufficient for his professional daily workflow.
He considers Claude Max a better value than Google AI Ultra, based on the usage, ecosystem, and models discussed in the video.
The host recommends giving agents ways to verify their own work, such as generating and running end-to-end tests.

Chapters

Start	Chapter	Summary
00:00	Argon’s launch and scope of the review	The host introduces Gemini 4 Argon and outlines a review of its access, cost, benchmarks, practical performance, and usage limits.
01:01	Availability, features, and pricing	The video covers Google’s public release, the model’s stated use cases, and its introductory API pricing.
05:29	Cost, speed, and Google’s use cases	The host compares model speed and token costs, then reviews examples of Google using Argon for engineering, science, and 3D generation.
07:08	Access through Google plans and Antigravity	The host explains where subscribers can use Argon and discusses the plan restrictions, student access, and Google’s CLI and app options.
13:04	Benchmarks and competitive position	The host compares Argon’s benchmark standing, task costs, and reasoning steps with competing models, concluding that it is in the Sonnet-level contest but behind Opus.
16:26	Why agents should verify their work	The host discusses self-verification through tests and introduces TestSprite as a tool for validating agent-generated work.
18:06	Testing Argon on a coding task	Using the Antigravity CLI, the host attempts to improve a 404 page and comments on permissions, latency, model questions, and the results.
28:17	Usage limits and final verdict	After reviewing the task’s quota consumption, the host argues that Argon is difficult to use as a daily professional model and compares Google’s plans with Claude’s.
32:03	Audience question and closing recommendations	The host asks viewers whether they would switch plans and closes by recommending self-verification and further study of agentic programming.

Quotes

now it is available to everyone — 01:01

Google’s models are very good. — 08:13

auto-approves edits and commands considered safe. — 21:52

It spent 10% of my weekly limit to do a single task — 28:17

it’s hard to use because your subscription would have to be much more expensive — 30:59

Concepts and entities

Gemini 4 Argon, Claude, Sonnet, Opus, GPT-6, Antigravity, TestSprite, Google, Anthropic, Rafael Quintanilha

Full transcript (original language)

00:00 And finally, Google has made Gemini 4 Argon available, both via CLI and through its artificial intelligence plans, provided you are on the Pro plan or higher. So, in today’s video, I want to tell you everything about this new model that has arrived to compete at the top, both with the ChatGPT family models, GPT-6 Astra and GPT-6.1 Sol, and with the Claude family models, Op- 5.5 and Sonnet 5.5. I want to walk you through the specifications of this

00:31 model and how much it costs to run. We will compare community opinions with what the benchmarks say. I want to perform practical tests and tell you if the limits are holding up, and answer the question: is Gemini 4 Argon the model to end the sovereignty of Op-5.5? Folks, a few weeks ago, Google had announced the launch of Gemini 4 Argon, which was indeed coming to compete at the top level, but we didn’t have

01:01 access to it; it was only for a small group of people due to security reasons and so on. Not anymore, now it is available to everyone, and that is exactly what I want to talk to you about in today’s video, alright? So let’s start by going through the official launch post, so to speak, right? The public launch, right? Which happened on October 13th, and Gemini 4 Argon is now globally available to developers and businesses, both through the API in AI Studio and the Gemini platform, as well as for Google AI

01:35 Ultra subscribers, and now Google AI Pro users too. There are some nuances regarding Google AI Pro subscribers. And if you are in the Google Plus subscriber group, well, then you won’t have access. We will get to that point in just a moment. You can then use it within the antigravity. Excellent, very good. We have a proprietary harness that we can use for Gemini 4 Argon, okay? So, what are the main features of this model? Well, it excels in complex

02:08 work, right? Work more focused on real-world tasks across various domains , right? Including programming, finance , legal, right? Cybersecurity and science. One thing that catches my attention here is the fact that they are highlighting programming. We can always tell from company releases which is the first highlight they emphasize, right? And here they are indeed highlighting the programming area, which is always interesting. The price of Gemini 1.5 Pro, it’s at a price they’re calling introductory, right,

02:42 which goes until November 30th, okay? So, until November 30th, there’s about a month and a half here for Pro to be at a price of 3.50 per million input tokens and 10.50 per million output tokens. So, basically, it’s coming in at the same price as GPT-4o. Claude 3.5 Sonnet, right? If we look here at OpenRouter, we can even compare these models quickly to see Pro alongside Sonnet and also GPT-4o, which

03:16 are models in roughly the same price range, okay? Actually, I think it’s worth it if you want to compare these models to watch two videos I’ve already recorded here on the channel. The first is a video where I say that the Claude plan, the Claude Pro subscription plan, got even better after the release of Claude 3.5 and the price reduction for Sonnet 3.5. I’ll leave the link here for you to watch later. And the other video that will be interesting for you to watch is a video

03:48 where I ask if the ChatGPT Plus plan, especially OpenAI’s, makes sense after the launch of GPT-4o. I’ll also leave it here for you to watch. In that video, I also talk a little bit about Pro, okay? At the time, it was when it had just been made available. If we take it here to compare within each provider, we have Gemini 1.5 Pro, Sonnet, and GPT-4o in the same price range, with Sonnet coming out, you know , as the cheapest in weighted average here, considering the tasks that OpenRouter uses. Usually, this happens

04:24 because the model needs fewer steps to reach the final result, and typically, Gemini models always needed many more to complete a task than a model like Anthropic’s models or even ChatGPT’s models. This has a big influence because the more the model needs to do, it’s as if it were traveling more paths, it’s like you’re spending more gasoline in the car, right? The gasoline for the model is the tokens it produces, let’s put it

04:54 that way. So it’s producing output tokens while it’s thinking and evaluating various probabilities, various paths, and this generates more tokens, making it more expensive. That’s why Gemini 1.5 Flash, despite being a faster and cheaper model on paper, didn’t end up being so, because the model took so many detours to reach the result that it used more tokens and ended up costing more. Here we see Gemini 1.5 Pro following this trend, though not quite as much, not quite as much. It’s about twice as expensive in terms of total weight here compared

05:29 to Sonnet and 50%more expensive than GPT-4o would be. So it’s right there in that range. I think it’s fair to say here that they are in the same ballpark, okay? Now, regarding TPS, there’s no way around it. Anthropic remains far superior to both the GPT-4o versions and the Gemini 1.5 Pro version . Even when we switch providers, from AI Studio to Vertex. Yeah, the TPS is 30, 30, 40, it’s quite low, while at Anthropic we can hit 95 tokens per second on the subscription. If you run it on the API, it goes up to 100, right

06:03 ? And on other providers, you can even exceed 100 TPS, okay? So, in fact, Anthropic models are currently unbeatable when it comes to speed. But getting back to the announcement, how does Google say they are using Gemini 1.5 Pro within Google itself? So, to solve engineering problems, right? Codebase refactoring, optimization, both of the codebase and business workflows within the company. Again, across various domains, right? Focusing

06:37 not only on programming, cybersecurity, etc. They have a set of things they’ve done, like Python codebase optimization, science-related tasks, quantum chemistry, and package optimization. They are also really pushing 3D generation, which we saw was something GPT-4o Astra brought well to the table, right? Very good 3D generation. Actually, I made a video about this too, I’ll put it here for you to watch later. So, in fact, what

07:08 Pro seems to do is match what models from both Anthropic and OpenAI were already doing, okay? It remains to be seen if it’s enough to make you use it as your daily driver, which is what we’ll see next, because Gemini 1.5 Pro is now available on Google AI Studio, okay? which is here, Harness, where Google is focusing its programming tasks. So, GMN 4 Argon can be used inside Antigravity there for Pro and Ultra subscribers, right?

07:40 Available in the Antigravity desktop app and also in the Antigravity CLI, which is what I want to test with you. Parenthesis here, I recently did a review of Google’s AI plans, where I even talk about the Antigravity CLI. So I’ll put the link for you to watch later, especially if you’re in doubt about whether it makes sense to subscribe to this AI plan. And now, having this information that you can get access to Gemini 4 Argon, which is their main model, okay? And again, here

08:13 they give several examples, eh, legacy code migration, YouTube comment optimization, ticket triage, nothing very new here, right? Doing this here, look, text localization, right? This is something very good, right? We’ll see here in the community opinions that Gemini models, and I share this opinion too, they tend to be the ones that express themselves best. So, if you need a task to produce text, write text in a more natural way, Google’s models are very good. So, I’m curious to see how Argon does in this regard too. All

08:46 right? So, OK, Argon is available for subscribers, right, of Google’s AI plan . They rolled out the model, right, after a certain time, right? So they started releasing it here to the people who are on the AI Pro plan. But there is a detail, eh, many people have this AI Pro plan through some promotion. For example, there are many people here in the X comments saying they paid for the Pixel, right? Bought a Pixel, Google’s phone there, precisely because they would get this plan for a certain time and these people are not actually

09:18 covered, right? What happens? If you look here on the Antigravity page itself, they say here, look, the Gemnight 4 Argon is available from Google AI Pro onwards, okay? So, Google AI Plus and those who are on the free plan won’t receive it. And from what I’ll show you here, it’s not in the plans, okay? Even because in the official release itself, right, post here by Logan K. Patry, who is the person mainly responsible for AI at Google, at least the guy who does this publicly, he said here, look, Argon is

09:48 a large model, right? So, we have limited availability for it, but we are working on versions of Gemnight 4 that will be available to billions of people . So, it seems to me that Gemnight 4 Argon as it is won’t be coming to the Plus plan, okay? So, if you are on the Plus plan, don’t expect Gemini at its maximum intelligence. Which, let’s face it, is somewhat understandable, because Google’s AI Plus plan, let’s even open it here, if we open the Gem Nike plans here, look, they are even already marketing Gemnight 4 Argon. The AI Plus

10:23 is R $ 24.99 per month, so R $ 25, we’re talking about 5 here, okay? We don’t have any subscription for $ where you get high usage, right, of such a capable model, okay? It usually starts at 20, which is more or less the AI Pro , it’s even a little cheaper here and still comes with all these things, like storage, right, and other benefits I showed in the video I mentioned to you. Then there is a detail that many people get the student discount here and then this price becomes very cheap. Then you might be wondering, will the student

10:55 discount have access to GMN 4 Argo? They say that Gemnight 4 Argo is only available to those not on a Trial, so you have to be paying something. If you are on the Google AI Pro plan and leaving some change for Google, you will have access. If you are on a completely free plan, for example, a Trial, or if you bought a Pixel, or even have a plan, if I’m not mistaken, an Indian one, right, that people have access to for free. If you are on that, you won’t have access. You have to be giving something to Google there, otherwise it won’t let you use it. Now, a curious thing too, right? You see

11:28 that there are these three asterisks here on Gemn 4 Argo, then you will see that it is not available in the Antigy IDE, so it’s only in the Antigravy app, the CLI, and the IDE extension. So this tells me that possibly we don’t have a very bright future for the IDE going forward either, okay? It seems to me it will be the app and the CLI, which let’s face it is also the same approach taken by Openi, which has the Codex Desktop app and the CLI, right? And also by Antropic, which has the cloud desktop app and also cloud code, okay?

12:02 So this is for you who came to the comments to ask me why I didn’t review the anti-heavy and it’s right here. Why ? Because I know Google has this reputation for killing off the products it launches without any warning. So, you know, it didn’t make sense and I really don’t think it does. CLI is different because with CLI, you can usually integrate it in other ways, and that’s what I’ll use to test in today’s video. I even left a question here on my Twitter for people who have the student discount to see if it was working, if they were getting access, right? And yes, right? A lot of people

12:32 told me here that the student discount is working right here. Philorock Degf too, Felipe Luiz, a lot of people saying that the student discount is working, but the limits are quite low, okay? The limits you can use within this student plan, which would be the pro plan, are quite low. So, we’re going to do this test in just a moment. Before we get to the practical part of this video, though, let’s see what to expect from this model, okay? Why? Because this was a model that came with a lot of expectations to compete at the

13:04 top. If we look here at the benchmarks, starting the theoretical analysis, let’s say, Gemini 4 Argon comes in here with the same score as GPT 6 Astra and Fable 5.1, right? Very interesting, still below Sonet and Opus 5. But right there, fighting with the new Mus Wonder , which actually came out this week. Also, if you want me to make a video about Mus, tell me in the comments. I’ve already made some videos on the channel. It seems like people in Brazil

13:34 here don’t have much interest in testing Mus. So let me know, if there are people who want it, I can make a video later. But it came in here, more or less ahead of GPT 6.1 Sol. So, if you have an OpenAI subscription and you’re possibly using GPT 6.1 Sol Plus because Astra consumes a lot of tokens, you’d have a model there that’s compatible with the OpenAI subscription , which is quite interesting, okay? And you can see there’s already a leap in quality here compared to another group here, which would be the Grock 4.7 group and other open-source models, okay? So, at least in the benchmark,

14:07 right, it came to compete in this little group of GPT, Muse, and the previous family of Tropic models, because the 5.5 family is really very good. So, we have good prospects regarding this model. Now, in cost per task, Gemini here sits between Sonnet, or actually well below Sonnet. Here Sonnet, in artificial analysis, is much more expensive than Argon and also much more expensive than GPT 6.1 Sol. So, GPT 6.1 Sol is truly the affordable model. Gemini Argon would be about three times more expensive here, but

14:42 still 40%of the price compared to what Sonnet 5 would be. Right? So it would fall between these two models. It is true that Sonnet 5.5 is ahead in terms of intelligence, right? It is well ahead here. Over on the Cursor benchmark, Gemini 4 Argon has joined the fray. As I record this video, it enters in sixth place. Look there, below Sonnet, but ahead of Opus 5. In reasoning medium. We don’t have tests with GPT 6.1 here in Cursor because it

15:17 wasn’t available in Cursor. But it’s interesting, see. It comes in here to compete with Sonnet again. Only here in Cursor, it’s turning out to be more expensive, okay? So you see it’s all kind of mixed up, right? Sonnet still seems like a very good value, guys. If you’re worried about price, Sonnet 5. It really is an excellent model and hard to beat. Regarding the steps, right? Gemini 4 is here with quite a few steps, but not too many, right? Not that many, right? We do have a distortion here from the max mode,

15:50 which always has more steps, but we could say it’s more or less in the range of Opus 5.5 extra high, right? Sonnet 5.5 extra high is already much more direct here. But anyway, the benchmarks confirm that Argon is in this fight, it’s in the mix, at least at the Sonnet 5.5 model level. It seems to be the consensus that it is below Opus 5.5, okay? But at least in the fight with Sonnet 5.5, it seems to be there. Now, regardless of the model you select, whether it’s Sonnet, Opus, GPT,

16:26 or Gemini 1.5 Pro, it’s important that you have a way to allow your model to self-verify, to validate the work it is doing. And one of the ways to do this is precisely through a CLI, like the TestSprite CLI. Guys, TestSprite is a partner here on our channel that has a very interesting CLI that provides a verification layer for your agent. If you do good planning and have a way to give your agent a way to verify itself, for example, by writing end-to-end

16:58 tests and ensuring it only returns to you when those tests are passing, you don’t need to keep chasing after the most expensive model, you know, the largest and most intelligent one possible, because it has validation mechanisms. Maybe it will take longer, but it will reach the result you want, okay? So, TestSprite is excellent for this, so that you have more confidence in the model you are using. Give it a task, ask it to generate end-to-end tests. With TestSprite, it will write these tests, it will validate itself

17:35 against these tests, and you will have the guarantee that when it returns to you, those tests will be passing, which greatly increases our confidence in working with agentic programming. Alright? If you want to test and get to know the TestSprite CLI for free, the link will be in both the description and here in the pinned comment. Very well, everyone. So let’s go to the practical part of our video today, where I will run the AgentGPT CLI. You see that I have the most up-to-date CLI possible? So I’m going to launch the

18:06 CLI here using the agent command. 1.5 Pro with reasoning set to high. How do you do that? You come here to model and it will appear, look. I have both the Pro and Flash, and even Opus. Here in AgentGPT, right? Pretty cool. So you can select the Pro here and you can also change the effort, okay? I set the effort to high, alright? I always like to use the effort on high, if my limits , you know, allow it. And speaking of limits, let’s go ahead and check our

18:38 quota here. And I have here, look, the quota for my Gemini Pro models. Alright ? So, you see that I have everything in the green, I use the AgentGPT CLI. Those who follow me here know this, so I really want to run these tests, okay? The first thing you’ll notice is that you’ll have separate limits for Gemini models that aren’t Argon, right? A limit for both weekly and 5-hour periods. And you have here, right, separate limits for Claude and GPT models. In this case, GPT here is the GPT OSS, right, which is their

19:09 open-source one. This one you likely won’t use. So, basically, it’s a limit you have to use Opus and Sonnet. This is pretty cool. And finally, what interests us for today’s video is the Gemini 1.5 Argon, the weekly limit for those on the Pro plan or above, with the quota being separate. And of course , if you are on the Ultra plan, you will have even more limits. So this is what I want to verify today, okay? Let’s check how much we’re going to spend, right, on the tests I perform. So what do I want to do with you today, folks? I haven’t prepared anything very specific. Actually, what I want to do

19:42 is fix some things I have in the AI program, right? The AI program is my AI engineering training; it’s already running with several classes coming out every day, and people are really enjoying it. And so I’m tweaking some small things here, okay? I don’t know, if I make up a page here today, it has a generic 404. Can we make a 404 that is closer to what the page’s visual identity would be? That’s a first thing we could test, okay? So, a simple UI task here. Let’s start with that one, then. So, for starters, I’ll ask it to

20:14 initialize my development environment here. Please init environment so locally. So I’m going to do several things here to mix in typical programming work. Why, folks? Because this is a good model, that’s undeniable , right? I know it will solve the things I give it, and I don’t like that type of benchmark. Create a game from scratch. That has its value, but it’s not what I like to bring here to the channel. I like to bring to the channel things for everyday use, okay? So, what I want to do here is simulate the daily

20:46 life of someone working with software, someone who works with programming agents, and see what the work experience is like. I’ll give you my personal opinion here, you may agree with me or not, and also see how the limits hold up to check if this pro plan is enough for those who work with this. So, the first thing I’ll mention to you is this Google part here , see? It doesn’t come with a very smart permissions system by default, okay? It’s asking me for authorization to read the readme. Why, in the middle of 2026, do I have to give authorization for my AI agent to

21:19 read a readme file? So this won’t work, I’m going to stop this, you can’t work like this, okay? I saw someone saying they had auto mode enabled, okay ? Let’s see if I can find something like that here, look. It’s project-based. I’ll do it globally, man . I want it global right here, look. No , if I go here, it will ask me to select the tools I want. I don’t want that. I don’t want that level of granularity. Let’s go to settings and see what we can find. We have agent mode, look. It’s already here, you see? There’s default, accept edits, and the mode, alright? So, auto-mode

21:52 auto-approves edits and commands considered safe. You can use a classifier for that. Most companies do it this way, and it will ask you when there’s an irreversible action, okay? So, of course, this should be the default, right? For heaven’s sake, I don’t understand why they don’t make this the default. But fine, I set it to auto mode and let’s go. Go on. If you don’t set it here, it will keep interrupting you every second to see if it can read the simplest file in the world. So, put it in auto mode, that’s the first tip here, otherwise you won’t be able to work with Antigrab. Look,

22:22 just like in the other video I made, it wants to create a new database. Why is that? The BD already exists. Please just start the server. Look guys, I tried, but even its auto mode is basically useless, okay? It’s asking for permission to run the server development command. So, I’m going to stop here, I’ll have to go back to config and I’ll have to set it here. Man, there’s not a single thing I can just do. Just go and run everything. Don’t ask me anything. I’m going to set it to always proceed. Here comes my main criticism, because this “always proceed” isn’t great. It keeps asking

22:55 me so many things that I end up accepting everything, so when something actually dangerous comes up, I’ve already accepted it because I’m conditioned to say yes and have to keep granting access to everything. So yeah, they need to improve this auto mode of theirs, okay? This part is bad. Another thing I wanted to talk to you about is the speed, okay? The first tests here aren’t very satisfactory, not at all. I find it quite slow just to run a simple command. So look, man, there are some things you just can’t understand. It was just supposed to run “pnpm dev” and it decided to do a database migration. This Google thing isn’t that great, but fine, it’s working. If I open it here,

23:30 okay, it’s loaded locally. So I’m going to ask it to create, you know, a 404 page for me, alright? Alr so you gna see that the 404 page that we have pretty simple and has nothing to do with the visual identity of the app. I wonder if we could design a better 404 page. So this is a task that is meant to be strictly frontend, okay? Right? Comparing it directly to the cloud code harness, I miss having the cloud code canvas. It’s really good nowadays,

24:03 right? I can create the design within the design system, see it there, and approve it before development, right? In this case, it’ll have to create it from scratch, right? It will have to create it while at least reading the design system. What I saw is that it already managed to do it, look, it already managed to reference the design MD. So, it had access to my design MD, to see how my design system is laid out . Let’s see how it’s going to build this page. Now, a tip here is to press Ctrl + O to see exactly what the model

24:33 is doing, right? We can see here all the files it’s loading, you know, into its context. This is interesting for you to correct the model if it’s heading in the wrong direction, which it seems like it’s looking at what actually needs to be done, right? Alright, guys. It took a little while here. It did some exploring and now it asked me a question. I found that interesting, okay? The question about how I want this page to be, since I was being very vague, right? So, which direction? The recommended direction is editorial, hero, wayfinding grid, navbar and footer monospace, blah blah

25:04 blah, okay? Minimal, editorial statement, generous negative space, technical spec card, I don’t know. I’m going to tell it something like: “Try something more in line with the vibe of the app”. I don’t know, I want it to be a bit creative here, right? Like a 404 page that’s a bit more creative, let’s put it that way, right? Yeah, it suggested a bunch of options here that honestly didn’t mean much to me, okay? So I’m going to let I’m using this here

25:36 to deliberately explore the creative aspect, right, of Gemini 4 Pro, particularly because I’m on high reasoning here, okay? Let’s see if it comes up with something. Guys, after a very long delay here, Gemini 4 Pro got back to me, look, the realization of what it did. It updated the 404 page using the not found component and the same layout here, okay? OK. Let’s go back there then. Let’s generate some random page here. And congratulations,

26:07 page not found. This page doesn’t exist in the address engine, OK? Right? Nothing special. I would like this 404 to be a little bit more highlighted here, okay? I thought it got too hidden here. Yeah, but that’s it, right? Nothing special. The task was simple, right? It took much longer than I would have liked, okay? It took much longer than I would have liked. So let me ask it for this this adjustment, okay? I’d like the 404 to be a little bit more highlighted. It’s very difficult to see this way. And also I saw you working on two apps. Work only on the Vite app for now. A detail that happened. I switched

26:43 the program code from my NextJS repository to Vite to be entirely on the Cloudflare stack. And I didn’t tell it that. So that wasn’t the model’s fault, it was my fault, okay? Since I didn’t tell it that, it worked on both projects, both NextJS and Vite. So, I need to be fair here. This certainly made it take longer, since it had to explore two different codebases. Still, I think it took longer than it should have, but fine. This this in fact had an influence here, okay? So let’s see what it what it comes up with now with

27:13 this new interaction of mine with it. Now, as the model works here, it’s important that you understand, guys, that the more leeway you give the model , right, to make decisions for you, the more prone you will be to having to make iterations. In fact, this is one of the topics I talk about a lot in my agentic programming course. If you go to programa.ai and head over to the agentic programming course, this is exactly one of the points I really hammer home, okay? Right? Why do vague prompts fail? So, if you want to avoid this kind of delay we’re having with

27:46 Gemini 1.5 Pro, and actually get the same kind of results with faster and cheaper models, consider checking out the course linked in both the description and the pinned comment. All right? Oh, it did what I wanted, reverted the change that wasn’t needed anymore, and only touched the app.tsx. Let’s open it up here now. Yeah, it added a little tag here, right? No, that wasn’t exactly what I had in mind, but it’s okay too, right? I didn’t give it anything, no reference at all. It just created something random here, okay? It used the design system. I can’t deny that. It’s working on mobile

28:17 . So, okay. Let’s consider this one done. And now a big thing, let’s take a look at the usage. And this just won’t do, right? This is truly the big problem with this model, the way it is right now, okay? It spent 10%of my weekly limit to do a single task, including what I would do day-to-day, like initializing the server, right? Running a prompt, asking it to adjust something else. It spent 10%, guys. You’re going to burn through, right, your entire weekly limit on this. For those who work with programming, this

28:50 will be gone in a day, okay? At most two days, this is gone. So what we can firmly say at the moment I’m recording this video is that the pro plan, right, the plan starting at $ 100 and up, or if you have a discount it might be cheaper, but their pro plan for professional work isn’t going to go very far, okay? You will have to be on the ultra plan if you intend to use Gemini 1.5 Pro as your daily driver. And so, guys, there’s no way, right, we can’t talk about the elephant in the room, which is the fact that we have subscriptions at the same price that offer much more usage, especially with Claude Pro, okay? If you subscribe to

29:25 Cloud Code today on the $ plan, you will get access to Sonnet, Haiku, and Opus 3.5. They are excellent models. Sonnet and Opus are above Gemini 1.5 Pro in the two benchmarks we saw, and they will perform much better. Even with Opus consuming more within the $ 20 plan, Sonnet will certainly see more use than Gemini 1.5 Pro. So, although this is interesting, I purposefully didn’t explore the maximum capacity of Gemini 1.5 Pro in this video. I know it

29:56 is a much more capable model than just making a 404 page. But what I wanted to show you is that, within a daily workflow where you perform multiple tasks—asking it to investigate, initializing a server, running commands to build a page, and iterating—is that in just a few minutes of use, besides the latency I wish were faster, the weekly consumption really isn’t suitable for those who want to work with this professionally. I repeat, especially if you are on the Pro plan or the Ultra plan, then you might

30:27 possibly have a much more interesting use case. But if we go to the Ultra plan, man, we’re talking about almost R $ 800. And if you compare that with a Claude Max plan for $, which ends up being between R $ 500 and R $ 600 depending on the conversion, man, it doesn’t seem like a good deal to me, okay? The Claude Max plan today, for what it delivers, the ecosystem, and the models, would really be much more worth it than Google AI Ultra here, okay? So, to summarize my video, what do I think of Gemini 1.5 Pro? It’s cool

30:59 , a model that came to compete, but in practice, for daily use, it’s hard to use because your subscription would have to be much more expensive for you to use it day-to-day. And for the price you pay for that subscription, you would be better off with another one, okay? So, if you have the AI Pro plan with a discount, paying like R $ 20 or R $ 25 to unlock Gemini 1.5 Pro, alright? Use it within the app or the CLI, maybe you can use it for one or two more complete tasks during your week, but know that it won’t be able to

31:32 be your daily driver, okay? Personally, I think it might be interesting to use it as a text editor, perhaps? You generate, you know, copy and text with your ChatGPT models, with Anthropic’s model, and pass it through Gemini 4 Argon so it reviews the text and makes it sound a bit more natural. That could be a use case I’ll explore in upcoming videos, alright? But now I want to know from you what you thought of Gemini 4 Argon. Do you think it’s worth it? Tell

32:03 me here in the comments, I want to know your opinion. Would you switch your OpenAI plan or your Anthropic plan for the Gemini plan with Gemini 4 Argon, or would you rather stick with those alternative plans? I want to know the community’s opinion. Since you’ve stayed with me until this point, subscribe to the channel here. A lot of people watch the videos but aren’t subscribed. It helps a lot if you subscribe, like, and share this video. If you’re on your phone, of course, give it a like, it really helps us out.

32:35 Don’t forget, a good way to make your limits go further and avoid rework is to give it a way to self-verify. And for that, I recommend you take a look at the Testprite series. And if you want to dive deeper into the study of agentic programming and how to use agents to develop software and apps, take a look at the agentic programming course over at Programa AI. I truly thank you for watching this far and see you next time.