Should You Use OpenAI's New Browser?

OpenAI just released another big thing. They have their own browser now and it's making real waves all across the internet. And I really wanted to make this video to talk about if this is any good and worth your attention. How this compares to some of the competition out there cuz this product doesn't exist in a vacuum because consumers have many choices these days. And three, how capable is it actually? So let's start that discussion by summarizing what this actually is. It's called Chachib Atlas, a brand new browser that is as of now available on Mac OS for all plans. Yes, that includes the free plan and there's no geographic restrictions here. So even people sitting in Europe like me can try this out right now. If you're on a Mac, simply download this, install this, login with your Chat GPT account, and off you go. This is what it looks like. It's basically a browser that has Chat GPT as the default new tab window with a few things that should be noted. and it brings some unique features compared to just using chatb in a tab as everybody has been so far. First of all, and kind of obviously if you're on any website, you can just click ask chat here and you have it right there and it has the context of the entire site loaded into the chat. No need to copy paste. You can kind of just talk to it. And at this point, I want to start pointing out how this compares to some of the competition cuz this is not a brand new idea. These Chrome extensions that do this have been around well since the advent of chat in 2022. It's a useful thing, but honestly, copying something and moving it to a new tab isn't that big of a pain point. So, call this one a nice to have. Secondly, and more interesting, so they have this concept of browser memories. So, if I go here to my settings and I go to personalization, you will see a new category here where it says reference browser memories. If you're familiar with chatdy memories, this is the same idea, but from your browsing experience. So, as it learns about your internet usage, it adds them here. This is a thing I don't dare to express an opinion on yet cuz time will show how this is actually implemented. And this is a very unique feature that I believe we haven't seen from any other product yet, but at the end of the day, it's just another way to automatically gather context and save it to your account. Beyond that, I should point out when you install this, it asks you to import all of your bookmarks and your history from Google Chrome. And this kind of launches me into the biggest question and concern here, which is, okay, so you install this and then you decide to give all of your browser history and all of your bookmarks to OpenAI just by importing it. Hm. Sure, that can be useful in some of the things, but just be aware that that's what you're doing when you're installing this, importing everything over. And yes, it will open up some nice opportunities, but you might not want to share your entire browser history with OpenAI. And this is kind of a sneaky way of them getting it. So, just be aware of that. What I could do now is something like I did in a previous chat here and ask something like find movies and shows I recently viewed. Scan my recent browsing and surface movies and shows I viewed or searched for recently and then use my browser history as context to correctly identify that yeah I checked out trailers for the bear pantheon the Gothic one remake game trailer. That's a good one. And then make suggestions for what I should view next which actually I've seen half of these. So this is quite good yet again none of this is really novel. We've seen this from other AI browsers and from Google themselves where you could use your Google Chrome history as context in the chat. And you might see the theme here already. It's sort of pulling these various things together. And that brings me to the next big point which is the agent mode in here. And this is the one where we'll start getting practical. And I want to actually start by running a rather complicated prompt for the agent mode to execute. It's going to be one that includes three different applications. Before I run this, I just want to quickly intro agent mode. It's basically their version of what would be called a computer use agent, an AI that has access to a computer, to a browser, and then it does things, clicks buttons, gets things done. Usually, they take the same demo of like booking travel, which is not a real painoint people have. I don't understand why they always use that as a demo, but that's what this is. It's basically agent mode that you already had in chat just within here. Now, it has all of the context, the history, maybe even the passwords that you have in your browser if you store them in there, which is just a more user-friendly way to use this. But basically, it's the same agent mode. It's a very interesting app. But here's a quick word from the sponsor of today's video. Okay, so one of the things I noticed with the new video generation models is that while they're amazing, most people don't actually want the new video generation model, they want one platform that can do it all rather than jumping back and forth between different tabs to make their vision come to life. And that's why I want to show you LTX Studio, the sponsor of today's video. So LTX Studio is a AI powered studio built for filmmakers, advertisers, and creative teams. It gives you access to cutting edge video and image generation models, but also everything else you need to take a project from concept to final cut all in one place. You can start with just a script, a concept, an image, or even a video. And LTX gives you professionalgrade tools to shape it however you want. You can generate production ready visuals, direct every shot with camera movement controls and key frames. It even gives you the possibility to guide the look with preset visual styles so your work stays consistent. There's also a feature called AI casting, which lets you create and reuse characters across scenes with custom wardrobes, so you don't lose continuity. And when your project is progressing and the time to polish it

comes around, you've got built-in tools like dynamic storyboards, a timeline editor, and even sound design. Again, all in one place. They also just launched a voiceover function, which is powered by Google Gemini 2. 5 Pro, and lets you use 30 professionally designed voices. Plus, if you're working with a team, all of this is collaborative. You can share projects, manage assets, and give feedback all inside the same workspace. So, if you want a platform that brings the power of AI video models together with the control and collaboration tools you actually need, check out LTX Studio. You can start exploring it today by using the link at the top of the video's description. A big thank you to LTX Studio for sponsoring this video. So, let's see if this works any better in this desktop version by trying out this test prompt. We're in step one, we're going to be reading in context from this document. I just copied a guide from our community over. This will be the context. The content here doesn't really matter. I just wanted to read it from that Google doc. Then I want to write an email with a summary of what the document includes that will be using my Gmail account that I'm logged into here. And then thirdly, I want to track all the emails that have been sent in this spreadsheet that is open over here. It's completely empty. And at this point, I'll just make sure agent mode is enabled and that I'm actually logged in with all the accounts that I have. Otherwise, these things wouldn't work. Okay, send this. And now let me tell you how this actually worked in the main competitor to this product which is Claude's Chrome extension. If we just briefly pull up my other browser, it's right here installed. And I actually keep using this thing ever since its release. And it's so damn good. It's better than all the competition. I know the top comment on the video when I made a video on that Chrome extension was, "Hey, what about comment? " Well, comet is good. It's the same concept as Atlas here, right? But it just doesn't work as reliably. And that's a big word in this discussion here. Reliability. Is this thing reliable? Because of course the idea of an agent that just controls your browser and does the work for you is great. But if you run a task 10 times and three times out of that it does something completely random, you're not going to end up using it cuz it's going to cause more work than it actually removes. And that brings me to the last part of this video, which is the various examples that I tested in here and how they contrast to some of the competing products as of the features. And I pointed out to you what it does and what competing products exist that do this already. Sure, this ties it all into one package, which is kind of nice. But then at the end of the day, if I want to get something done like the task we just launched, it's a real question if I should be using this cloud's browser extension, which has the same agentic capabilities, or something like Google's computer use model that they just released two weeks ago that has beat all the benchmarks. And let me just tell you on these agentic tasks, I find claude to be straight up better. And I don't think that's because the extension has some genius architecture running in the background or that chatbt atlas is just not good enough. I think it's because of the model. Brief reminder that OpenAI did not ship a new model here. They shipped a new interface that contains and uses the model they already had. And I'm really curious to see how this task actually performs because in the case of Claude that uses the Sonnet 4. 5 model where one of the big headline benchmarks and announcement was that it's so much better at computer use. It was so smart at doing this. If you saw that original video, you might remember that it actually did this task. It sent the email and at this step right here where I'm at right now, by the way, the computer is doing this, not me, right? It put in the summary of the email. It put in all the data correctly, but it was lacking a header. And then it actually went back and saw that there's no header and it deleted it, added the head header, and then moved everything down. Here in this case, it seemed to have got this right on the first try. So, that is nice. And that's actually it. It just finished up the task. So, let's see. It created the draft of the email. It's right here. And it saved the data that I asked for in here. And I remember this prompt actually just asks for the date and the email and not the content of the email. So it did what I told it correctly. But if it does this, it could easily also add the content over here that it wrote here. So on this task, it did equally as well as clot. Nice. Let me give you a few more examples that I ran before making this video on which it wasn't equal. And that should give you enough of a feeling for how this works. And that should round out the video and give you enough information to make a decision for yourself if you want to give this a shot yourself on your own machine. So, first of all, I just gave it a picture of Conor McGregor that was randomly on my desktop. I think for AI news you can use. I was doing some AI remix of it or whatever. And I asked it to generate a 100 social media captions using Claude AI for this image and put them into this Google sheet. So, what I found is that it actually did that, but it didn't use Claude in the process. So, it just opened up the sheet and it successfully put them in here. And look at these captions. They always like repeat four times, three times, just emojis or the hashtag varies. And it's good. You know, they're solid captions. Then I followed up and said, "Do 100 more by using Claude AI. " And then I provided the URL. And this would be my tip. If you're using this, give it the URLs and then it will reliably use the tools. What it did then is it opened up Claude, which I was logged into. It wrote this prompt right here, which I thought it was funny that it mentioned avoid mentioning the fighter's name. That's obviously the chat GP restrictions at work here. And then it just wrote 100 captions here. The formatting is a bit off. Doesn't really matter. It copied all over that. made one mistake where it just kind of copy pasted everything into one field

but then sort of it corrected itself and split them apart and there you go, another 100 captions from Claude. Quick note that I cannot withhold. I mean, look at how much more variety you get in the Claude captions. This is why Claude is so damn good for writing and why people really prefer it in many cases. You just see that right here. But we're judging chat GPT agent in chat GPT Atlas right here. It did this super well. So you can automate the usage of various AIs and the manipulation of data with spreadsheets, which I think is one of the main things you will want to do with tools like these, these computer use agents. It's like using AIS, researching stuff and manipulating spreadsheets, and that's what it's really good at. This worked well, but my tip would be very specific with the URLs of the tools and something like the Claude Chrome extension. It just pulled this up without the URL here. It needs it. Not a big deal. Then I did one more and this is a tricky one. We had our quarterly town hall in the community yesterday and at the end of it we just got together and tried this all together and we had multiple AI educators and university professors in the room and they brought up one topic and that is that these tools basically render things like LinkedIn learning certifications useless because the agent can just go through it, take all the quizzes, watch all the videos, pretend like somebody's taking the course and just farm certifications for you. So we put this to the test. We actually made it take one of the quizzes in our community concretely the final quiz on the advanced LLM prompting course and the results were very interesting because I gave the same question to both the claude chrome extension and to chatpt atlas here and I told it complete this quiz think thoroughly I really need 100% at all cost and here's what happened chatbt took about 30 minutes and it still wasn't done its method of getting all the questions was basically scrolling through all of this multiple times compiling the screenshots and then very slowly answering these questions. Claude did the same, but it must have been like 10 or 15 times faster than Chat GBT here at doing it. And both of them got 95%. And at this point, I have to actually admit my fault. I'm taking full responsibility for this. There was a mistake in one of the quizzes and that's the one question both of them got wrong. This question that is taught in the course, both AIs actually answered correctly, but we in our system had a wrong answer selected. So, long story short, basically both of the systems got a perfect 23 out of 23, but there was a mistake on our sides, which really tripped GT5 up. Kept running in loops after it was done and just couldn't finish the task because I asked for 100% and Claude was just like, "Hey, I did it and this seems right to me. " But even without those loops, Claude was around 10 times faster at actually completing the quiz the first time. So, yeah, that should make you think about all types of certifications that people show off on their LinkedIn. All you need is a free chat account in the Atlas to kind of do that. and it's a real mess for e-learning. So overall, what's the verdict here? Well, it's just OpenAI's version of something that we already had. If I had to pick one winner, which one of these computer use agents works the best? Well, it would still be the cloth browser extension with Sonnet 4. 5. That model is just the fastest and the most reliable at all of these tasks that I have tried so far. GBT5, this would be a close second though, and maybe this browser interface is actually preferable to you. I know a lot of people used Atlas before and really enjoyed that. But I personally think that a lot of these features like, you know, summarizing a website with like one click are a nice to have, but they're rather niche. It's not something you're going to be using on every site that you're on. And honestly, even if I had to copy this over to a new tab and summarize it here, it's not big enough of a pain point to change my personal behavior and approach to browsers. And above all, there's a real privacy concern where you're sharing all of your browsing data with OpenAI if you use this and even if you import your Chrome profile. So think about that before you try this. And lastly, I want to point out that this is just OpenAI's version of the open-source Chromium, which Google Chrome is based on. So it's nearly identical to Chrome, but it's set up to be AI first with interesting features that will evolve over time like browser memories and also you can do custom instructions for your agent mode here. Although I personally cannot edit these right now. So I think this is the first version that for the curious is worth exploring, but I expect this to get a whole lot better and featurerich over time. And I also expect them to do particular models that are really good at these type of tasks. Interesting release, not a game changer. All right, that's pretty much everything I have for today. My name is Igor and I hope you have a wonderful