I Paid 200 $ for the first ChatGPT Agent. Does It Actually Work? (OpenAI Operator)

All right, so it happened, the spontaneous Release out of OpenAI with the product that I personally have been waiting for the most in a long time. And I know a lot of people share these feelings because this is the agentic chatgpt version that so many people have been excited for, have been waiting for. I personally see this as the future of ChatGPT and really the direction that they're going to be developing everything towards. For now, it's a future that is only available to a selected few people, namely people in the US on the $200 pro plan. But no worries. Within a few months, they promise to bring this to all team subscriptions. And I believe that eventually, with all the competitors, just like we saw with the thinking models, I think even before that, we're going to get Operator products out of other AI labs that work at a fraction of the price. They're just the very first one to do it. But that is an exciting opportunity to actually give this a spin. In this video, we'll be going through this and we'll be running two of our very first operations within this new. You just access it like so. I am connected to a VPN to the US because I'm sitting in Portugal and I am on a pro plan right here. So again, those are the two prerequisites. If you match those, then you can use Operator for yourself today. Now, there's a few things I want to point out here that I think are interesting. Like you can schedule some of these and you can run free tasks at the same time. And there's some competitor products that I want to sort of compare this to. But we're going to do all of that after we run our initial operations, because I think I could be explaining this for a few minutes, but the best thing we can do is just show you, because what this does is it remote controls your mouse and keyboard. And that's what we're just going to do right here. Okay, first example, we're going to go easy on it. I'm going to say Airbnb, and as you can see, it comes with these partnered apps. There's more of them. I'll show you that in a second. But if you pick that, it's going to be using the Airbnb site that it has been like pre trained on, and it works especially well on it, so to say. And I'm going to be booking night in some place with Seaview in Lisbon. How about that? So book one night stay for two people in Lisbon for the. For January 24th, oops and yeah, we're going to have a look at this in practice. So this, this should be really interesting. I think the last thing that I might add here is maybe some filtration criteria. So maybe I'll add with C view and then I'll say under $300, I guess. Okay, let's send this. Let's see how it operates. This is going to be unedited because I want to really see and show you how this performs on the first try. Full disclaimer. I gave this one quick shot before I begin just to see if it's all working, if it's operating, I didn't let it finish. So this is literally a first look and unfiltered one. And at that. So a few more interesting facts here because one of my favorite things that they included here is actually its ability to save tasks. Okay, so if you have something that you want to be doing every week, like ordering groceries in a specific way, you can, you can save that and basically it can run once a week at the same time or you get a little preset for it. Okay, we can do that in a second. Show you that secondly, you can run three of these operations at the same time. So if I just go to create a new one, maybe let's do that right away. And how about this one? The second one we're going to do more Freeform. Okay. This used the Airbnb app that it had. But because Operator is using browser in the cloud essentially and just clicking, you know, buttons and using the keyboard for you, we can actually give it a task with any website here. Okay, so let me do that. Let me formulate a little prompt with something that I've actually recently done. I thought of this in advance. I think we're going to do reserve table for to people at Bernard. I think this is what it's called Esperanza. I think it's named like this in Lisbon for January 25, let's say through the fork. That's a website that I used to reserve a table there recently. And again, I could run this, as you can see here in the top left, it's running this one operation. Now it's running two operations. Operations at the same time. It's like cloning yourself now, like literally this one, this other instance of it is running and trying to book me the Airbnb accommodation while this one is running a search to reserve a table. So let's have a look at this complete Freeform browser use. You see my hands over here? I'm not doing Anything. This is just Operator doing its thing. And if this works correctly, it should change the location to Lisbon, and then it should pick the correct restaurant and book it. And I think, yeah, and we can watch it do that. Maybe we can open up a second tab, see how it's going with Airbnb. So this is really interesting because this is just the first step into this direction. And at this point, I want to point out that, yes, we've seen competitors do similar things here. Okay, this is not like a first. It's not like they, like, invented this. But if you look at some of the statistics, some of the benchmarks that I actually found super refreshing here, because these are just different benchmarks. I believe they were somewhere here. Or were they in the launch video as opposed to were in the launch video? This thing looks like. Look, it looks like this thing performs head and shoulders above all of the competition. And let's see if the practical application actually proves that, because really, that's all I care about here. Benchmarks. All right, fine. But will this get it right? Let me tell you. I spent about maybe 15, maybe 20 hours with anthropic's computer use, which is essentially this from Anthropic. Okay? It's a Linux machine. You install it through Docker, and then you basically have like a virtual computer that Anthropic's cloth can remote control that thing. Could not get one thing done reliably. Okay, I'm serious. Like, there was not one task that you could run twice that I was able to find that actually got the job done. If one of these two tasks work, that's already better than computer use, I can tell you that. Okay, so it was an interesting experiment from Anthropic, but in my books, not really usable. And I don't know anybody that has been using it successfully. So let's see if Operator is different here. What I really care about is if this works. So look, location, Waterfront. That's good. That's good. It's asking me a question. Okay, so should I proceed with applying these filters to view the available options? Yes. So as you can see, this is not fully autonomous yet, but this one is still running. And you can see all the steps happening here. You could also rewind, see how it was doing, what it clicked when it scrolled, all of this, but it will just jump back to the live view, where I could also take control here. Right. This is something they showed in their livestream. I could let it stop and take manual control, but I want to just let this run autonomously okay, this is looking good. And what about the second operation with Airbnb here? Also looking good. Look at that. Okay, there's some translation pop up. Let's see if it can deal with that. Yep, it closes successfully. This does have a sea view. Okay. You can see the ocean here. Well, to be fair, it's the Lisbon river, but it's sort of considered the sea and also like they advertise it at sea. So this counts. This works. So now it asks me a question. I found a listing in Lisbon with a Riverview that fits. It even says Riverview. Aha. Okay, smart, smart. Okay, it's a two bedroom apartment with a balcony overlooking the Tejo river, priced at 74. Okay, so with taxes and cleaning fee, 153. Should I proceed with the booking? Let me just say yes. That looks really good. Okay, so this is what I expected. The one with the application where it has direct access to Airbnb worked. Right? It's already out here booking it found a place successfully. I think this is the one where I'm curious. It requires an email address to proceed with the reservation. Okay, let me see. What if we did like a thing here where like, I don't mind, I can, you know, so I don't have a fork login. So I guess what I could do here, I could just click take control. Okay. It's operator browser. It's not screenshotting and it's all private. What if I actually went here and let me just quickly switch to the cam, log in with my Google account as I would. Right. And then let's see if it actually works. So, okay, successfully pasted my email here. We'll just have to have a few seconds of patience with me here. But luckily I had the login information saved. Saved over here. It wants two step verification from my phone from the YouTube app. Let me show you. No problem. Okay. Yeah, I'm in the US now. Google. No worries. Okay. And I confirmed it. And now I logged in manually. Okay. And now. Well, there you go, Igor P. This is my profile and I could basically say finish up and say finish. Finish the booking. Rerun control to operator. Let's see. This is the operation, by the way, that I literally did myself less than a week ago. So, you know, I could save a bunch of tasks like this in my favorite restaurants with the login information as a part of the prompt. And you know, if I have a burner Google account like that, I don't mind giving ChatGPT operator my password in the prompt preset. It could do all of this by itself and just book my restaurants for me. This is amazing. So while it does it, I just briefly want to touch on the model that this is possible with, because this is not good old ChatGPT4O. It's essentially a specialized version from a new model called Computer Using Agent, which is essentially GPT4O with vision that is trained on a bunch of computer usage. Okay, so they showed it a bunch of examples of people interacting and doing these tasks. That's why it works really well with things like Airbnb, because they just showed it, you know, thousands, millions of examples of people using their BNB website. That's why it works really reliably. And it has, you know, it's one of the preset apps. But from what I can see, it successfully picked. What did I want? January 25th, the fork. Two people like, this is looking good. It's still working. It's still working. But the only thing I had to do is log in with Google and this brand new. What is it called? Computer Using Agent model. Got this right? Let me tell you, there's no way on earth that any of the competing products would have gotten two requests like this right in a row. And from everything I'm seeing, this is looking very, very good so far here. I'm basically at the checkout. So I just, you know, I need to take control and log in with my Airbnb account, which I'm not gonna do right now because we did already one login. I showed you how that works. And the second one is still working. Let's see, what is it doing? Reloading page to finalize. Refreshing page. Like, this is very promising. Everything else I've seen, not so promising. Okay, and on the benchmarks, it just performs like 50% higher than all the competition on these agentic benchmarks. They only have two agentic benchmarks so far. But ladies and gentlemen, let me tell you, this is like the start of a new era. Like, I know I have a tendency to overhype some of these things and I get like excited by them. But like, this is literally, this is literally what a lot of people have in their mind when they think of AI. They don't think of this assistant where you have to precisely communicate what you need and then like follow up prompts to get some info on it. They think of like, hey, buy my groceries, book my restaurants, you know, do my work for me and take the work off me. That's what people mean. That's what people come in when they open that's the expectation they come in with. And right here it's doing that successfully. Isn't this unbelievable? Yeah, I mean, this is. This is great. Look, this is some, you know, restaurant booking site that is not a part of OpenAI's program. It's in Lisbon. That they didn't train it on specifically. Right. It had cookies in the beginning, I bet. Right. And it just works. So I could just take control here and confirm, or let's try and do this programmatically actually, as the magnum opus of this video. Let's just say confirm, return control to operator and let's see if we can book this table. I can let you know in the next video how the dinner went there on Saturday. But thinking, finalizing reservation, click confirm, your booking is confirmed. Wow. Okay, so that worked. Very first try worked. Second try with their app worked. What more can I say? So guys, quite unbelievable in my opinion, because I've played a lot with these products and they always just fail and it's always just a pain. Okay, but this worked. Now I could take this task and if I like going to this restaurant, I could just say save task and I could say save. This sort of like a prompt preset. This is a genetic preset. So I could save all my favorite restaurants. I could give it the login information here. Right. I could even give it the precise URL to the booking page of it to make it even faster. In this case, I'm just going to leave it like. So I'm going to say save. And then here we are, save tasks. And if I open up a brand new dialog, I have my custom saved task down here. And like, I don't want to overhype this really, but like, I booked this restaurant regularly and do you think the next time I'll be booking this restaurant, I'll just be. I'll be doing this manually. Heck no. We'll just go to my ChatGPT, press this button and then, you know, go do something and come back and confirm. That's way easier. Like, I think ChatGPT was so successful because for the use cases that really apply to all of humanity, it just kills it. It's like, you know, improve the grammar of my text, improve the structure, give me some suggestions, change the tone to this. Like all of those basic writing use cases, they're just such no brainers. Or like, here's a few keywords or here's what I like. Here's my rough email draft. Now turn it into something polished. That stuff helps so much and just applies to everybody. I'm not talking about these like super limited use cases that apply to a few people. ChatGPT got its success because everybody can find a use case for it that works. And I believe that if everybody had access to this for free at this point in time, I think this is a superior way to booking a table at this concrete restaurant than going through the whole process and Google login myself. And look, I don't mind giving it my Google login info on this account. It's just a burner Google account and I think this is promising. Now let's talk about the future of this to round it out. Okay. Because as you can see there's a lot of options here. We can look through some of these and I think there's also a lot of custom options. So at this point I would say if you want to run something concrete and Operator and you're not in the US and you don't have an account in the next 24 hours from upload, I'll be going through the comments here and I'll be running your guys prompt. So if you want me to see if a prompt works and if it succeeds leave a comment below. I'll take it, I throw it into here and then I'll reply with the result if it worked or if it got stuck. Okay, I think that's a good way to do it. Other than that, look, it can like aggregate news. I guess there was only one thing that I would like to correct that I said in this video, which is I believe it can save these presets for you, but it cannot automatically run them. And yet right in ChatGPT we got the tasks where you can schedule things up here and then they run regularly. It's just a question of time until operator gets this. Obviously the roadmap for this will be very interesting because there's many more things we can get. So to round up my thoughts on this, look, this clearly performs better than anything we've seen in this category up until now. This is clearly, as they stated, just a research preview that is going out to the smallest group. They have the pro users. Once it ships to the team users, I expect there'll be two more presets, more partnered apps, more of maybe you know, the little bugs that might appear. If I run 50 examples for this and not two will be ironed out, there will be more integrations. I can totally see how you will be eventually be able to save your login information to like your Google Account and your Airbnb account within ChatGPT they'll find some secure solution for that, I'm sure. And then on the other hand there's going to be an entire open source movement of people following OpenAI's lead just like they did with 01. Like we're gonna upload a music and use episode Friday tomorrow and like basically, you know, the Chinese copied like the idea behind 01 and did an open source version that is like 100 times cheaper now that you can use locally. I expect the same to happen for this. It will just take some time. So it's very early, it's very unique. I think my intuition on this, this is more useful than all of like AI video and all of, you know, GPTs and actions and whatever else we've seen come out of these companies over the past year or two for the broad masses. I'm not saying there's no use cases for that for certain individuals like in video production, but for the masses I think think that this is the most useful thing since ChatGPT and I think that's a big deal. It might take some time until this trickles down to Everybody. Right now, $200, probably not worth this, you know, operator and assistant service for most people. Nevertheless, I'm super excited. If you want me to test out your use case, leave a comment below and I'll spend the next few days looking into this deeply. We'll research it, we'll compile all the use cases that we find and I'm sure there's some interesting things you can do here beyond booking a table. And I honestly this makes me want to play with this stuff and yeah, should be an exciting future. This is the worst that will ever be and it's already working. Okay, so thank you so much for watching. This was an UNEDITED video on OpenAI operator First Look. What a time to be alive. Really like I think this is the sweet spot where AI really does make our life easier. What's coming in the next few years, I don't know. But right now I'm here to use this and to teach you how to use it yourself. See you soon.