# GPT 5.2 Is Here And I Tried Every New Feature

## Метаданные

- **Канал:** Skill Leap AI
- **YouTube:** https://www.youtube.com/watch?v=iIXRZkTWj8g
- **Дата:** 12.12.2025
- **Длительность:** 15:28
- **Просмотры:** 77,362
- **Источник:** https://ekstraktznaniy.ru/video/11956

## Описание

Here is the link to the official OpenAI post on ChatGPT 5.2 release: https://openai.com/index/introducing-gpt-5-2/

I go step by step to show how I tested ChatGPT 5.2 and compared it to ChatGPT 5.1 using real prompts for content creation, coding, research, and automation. I test the instant model, the thinking model, and the pro model to see how each version of ChatGPT 5.2 handles tasks like building websites, writing blog posts, creating slideshows, and reading images. 

I also test how well ChatGPT 5.2 remembers context, reduces hallucinations, and follows detailed instructions. If you're looking to see what’s new in ChatGPT 5.2 and how it stacks up against 5.1, this walkthrough covers it.

👉 Join the fastest-growing AI education platform! Try it free and explore 20+ top-rated courses in AI included Introduction to AI Agents course: https://bit.ly/skill-leap

## Транскрипт

### Segment 1 (00:00 - 05:00) []

Chat GPT 5. 2 just came out and in literally a single prompt created this entire slideshow for me with all this detail here. So, it pulled all this in from the sources and the web link that I gave it. I'm going to show you the three different models they just rolled out and on paper it looks really good, but we'll test it on my account here and I'll compare it to some of the outputs we got out of chat GPT 5. 12. So, the latest model from OpenAI is GPT 5. 2 and it comes in three versions just like 5. 1 did. It comes in the instant version, the thinking version, and the pro version. I'll show you what accounts and what paid subscriptions you need for those in a second here. And as usual on the OpenAI page, it looks like it destroys the old models in all the benchmarks. The thinking model especially is a huge leap over 5. 1. But again, let's go ahead and test it for ourselves. These benchmarks sometimes tend to be a lot better than the actual models. Now, it's supposed to be especially good at creating things like this is the old spreadsheet that GPT 5. 1 could create. And the new version with 5. 2 looks a whole lot better. And these are the type of tasks you probably do. I do this type of task all the time. And the output is never really good enough to use, right? You still need to do ton of manual work to get it to work. So, we'll test this out. Now, a couple other things I was really impressed with is they claim now that chat GPT 5. 2 thinking model will remember a lot in a single conversation. So, the context window is still the same. It's the same size. It's actually 256K token context window. Same as the other model, but in reality, that was not always the case. You would still forget things during a conversation. and it looks like this version of it now. It's almost at 100% the entire time through that context window. The vision capabilities, basically the ability to see inside of any image is also a big improvement here. And I use this all the time because I take screenshots of apps I don't know how to use and I ask it where do I click, how do I do this, that. Well, it's now able to see screenshots a lot more clearly and with a lot more accuracy. Now, this part's my favorite part if true. And this is my biggest problem with chatt probably yours too. The fact that chat GPT and every AI model basically hallucinates and makes things up. Well, right here they said chat GPT 5. 2 thinking now has a 30% reduction in hallucination. So GPT 5. 1 was at 8. 8%. GPT 5. 2 thinking is at 6. 2%. A 30% reduction is actually a big deal, right? If we're going to get down every couple points here every time we get a new model, well, we'll be at maybe 1% or lower accuracy, which will make it a massively, massively useful tool now for just about everyone. This is probably my biggest frustration when using AI since day one, but it has improved with every model release. Okay, let's jump in and I want to show you a couple of different examples they had. Then I already made a video about 5. 1 and I tested a bunch of different prompts there. So I want to show you those type of prompts in here so we could compare 5. 1 against 5. 2 really clearly with the exact same prompt and the same demonstration. The older models are still available in the drop down under the legacy model. So if you want to test this for yourself too, you can still use 5. 1 and all these other models over here. I'm in the pro account, so I have a lot more models than you probably will, but you still should have all the 5. 1 and five models. And the different models that come with this new 5. 2 release are three. One of them is called instant. This is when it doesn't think at all. That is not the thinking model. Then there's the thinking model, which for most people that watch my videos, I usually recommend you either choose auto, which decides between these two, or you on purpose choose the thinking model. It's a little bit slower, but for just about everything, even writing a simple email, I find it to be worth the few seconds wait here. And then there's the pro model. Now, the pro model is actually available in the chat GPT pro accounts, and it's also available in the business accounts, but it's not available in any of the other accounts like the plus accounts or the free account. You're not going to see the pro model at all. and the option where it creates really nice looking spreadsheets and things like that. That is also only available to the paid subscription, not in the free one. Okay, let's start with a prompt that they actually used as an example here. And this is a single page HTML. And if you use the thinking model and not the auto or instant, you'll see this option over here too. So I have this in my pro account where you could use different types of thinking. So in this case, I'm going to just leave it on the standard thinking here and let's see what we get. Okay, so it created the app for us. And the wind speed should affect what the

### Segment 2 (05:00 - 10:00) [5:00]

water looks like, the weight height. I could also alter that. And this one is the time of day here. That works. If it's stormy, does it affect things here? H some of it is working, but not working. And it looks like, yeah, I could look around here. There's the sun. Okay, not bad. I want to try the exact same prompt with GPT 5. 1 thinking model. Okay. And here is the one I got out of 5. 1. Obviously, it looks a lot more cartoony. Technically did what I asked it to do, but the other one looked like a real scientific test. This looks like some kind of weird animation I would have maybe as like a game or something, but not at all the same output out of these two models. Okay. The other thing that Chachi PT 5. 2 is supposed to be good at is actually creating things. So creating any kind of a website or a web app without you having to write any type of code here inside of chat GPT. So I asked it basically to create me a modern looking website and I told it to use the canvas mode which you can turn on here by yourself. This canvas mode lets you preview code in a visual style but I want to see if it follows my exact prompt there. And I give a ton of different details on exactly what I want. This is a website to compare different AI tools like we have on Futuripedia. Let's send this out. Okay, so this was the result that I got out of chat GPT 5. 1 thinking model when I ran this exact same prompt there. And we're going to look at 5. 2 now. Okay, at first glance, this looks great. The visual improvements are huge here. Light mode also works exactly as it should. Now, as far as comparing different AI tools, so far it's a little bit confusing and it's a lot broken, right? So, if I choose this model, okay, that looks good. The pop-up looks really nice. This transparent effect is also really good. So, I'm going to let's see. Select chat GPT to compare. Add a Okay. And then I'm going to compare it to Claude. Select to compare. And yeah, I don't know what's going on. So, it's not great. I I'm really a little confused about this formatting. Okay, I guess I could do that compare. Okay, that part works. This comparison works. But the initial app is kind of broken, right? The whole point was have a nice filtering system, but it's overly complicated here. And this part is clearly broken. So, again, I want to do this with just one prompt. I could obviously fix this with probably two or three different follow-ups here in chat, but the whole point is with one shot, I want to see if I could get these. And, you know, 5. 2 didn't really even hold up to 5. 1 with this specific test. The other one wrote maybe like 300 lines. Look at this one. I mean, this is wild. It wrote over a,000. Wow. 1,800 lines here to get us that app. So it's writing a lot more when you're creating any type of a user interface, any type of website, any elements of a website, which I use chatpt all the time for that. Cloud 4. 5 is actually unbelievable at that. So far this test doesn't even compare to that, but I'll save that for my ultimate comparison test. Now for creating any type of spreadsheets, let's create a project manager. Again, this is a prompt that they created and it's really elaborate here with a lot of little details. So, can this model follow these details and get things exactly right? Okay, the presentation is done, but it did take 28 minutes. So, in 28 minutes, it went through the link I gave it, different sources and things like that. But let me show you what it came up with because this is really impressive. So already chat GPT was never able to create something this nice. This kind of a layout was just not possible. But if you look at each slide, it probably is a little bit too detailed here. But the layout overall is shockingly good compared to where we were with 5. 1. It just wasn't able to do that. Now I use Gamma to create bunch of my presentations and it's really designed to do that. So, I really wasn't looking for a tool for it, but this did a shockingly good job. And you could also download it up here. And it downloaded it as a PowerPoint presentation, which then you could use the AI tools they have in PowerPoint to improve it, but it's awesome that it's able to do this. Now, let's do a quick writing test. Is it obviously better than the older model? Now, this I'm going to know a lot more in a week or so. Right now, this is the first day that this came out, but let's just ask it to write a hook for this very video. Now, I'm purposely not going to give it a lot of context, just the fact that I'm making a chat GPT video and I wanted to write the hook for me, but I specifically don't want to give it any context to see how well it looks in my memory and my previous conversation and things like that. Okay, so this is 5. 1 again. I chose the instant model so it

### Segment 3 (10:00 - 15:00) [10:00]

doesn't think 5. 2 just dropped and it's not an upgrade. It's a whole new era. Yeah, I would never say that. I thought AI couldn't surprise me any nope. Chat GPT is officially here and brings the biggest jump in speed. Again, none of these I would say. So that's 5. 1. Let's go to 5. 2 instance. Same prompt. This is not just another chat update. Chat PT 5. 2 quality change. Nope. Most people think chat 5. 2 is just faster. Nope. Okay, this one's not bad. I've tested chat PT 5. 2 across real workflows, content creation, research, automation, and this is the first update that actually changed my daily process. That's a lot closer to what I said. But again, it gave me bunch of them that I would never say. And on this account, I've actually tried to train it to follow my writing style, but I have a custom GPT specifically to do this type of hook. And I was testing to see how well it knows me here, not just how well it follows a simple writing prompt, too. Okay, this time I'm going to leave it on auto. This is something that a lot of models got wrong here. So I had this image and it needed to figure out which is the top view and all it had to do is just analyze these colors and see how they laid out in these shapes here. So again this does not require any prompt because the prompt is in the image. So it should be able to just read that and give me an answer. And it just thought for two seconds actually did not think for long at all. And let's see. Nope, that is not correct. So it got that wrong. Just for the sake of it, let's go ahead and try the same one with a thinking model. Okay, right away it got C when I turned on thinking, but it had to think for two minutes. When I had it turned to auto here and it decided between instant and thinking, it thought for two seconds and gave me the wrong answer. So, this is why I was saying even with 5. 1, I usually did not leave it on auto. Even though they're trying to go in that direction and let the model pick for you, it just doesn't pick right every single time. So in this case, it shouldn't have given me an answer in two seconds. It should have taken two minutes and give me the correct answer. I'll wait two minutes every single time. When I'm using this for work, I want the right answer. I don't mind the weight. In almost every single case where I use chat GPT in simple search, sure, if it's just extracting things with a search tool and I'm using it in replace of Google search, okay, that is fine. But in this case, I want the correct answer, not the fast one. But it did work. Okay, let's do a quick hallucination test to see if it's going to give us an answer. Give me the exact citation from a research paper where Albert Einstein first used the phrase black hole. Okay, it got that one right. It didn't fall for my trick. Einstein did not use the phrase black hole. He actually was coined in 1968 by someone else. So, he passed that part. But a lot of models actually get this type of question wrong because they just try to give you an answer and they just kind of make one up. But if you ask for a citation, a lot of times that does clear that up. Okay, this is something that I actually use all the time. And even though these models don't really understand word counts, I'm waiting for a day where they finally train it in a way where it does understand it and it could produce us exactly the word count we need. only because a lot of times when I'm writing a blog post, the title, the meta description, a lot of different things need an exact word count that I like to hit. So, let's see if you could write exactly a 300word product description for iPhone 17. Oh, wow. Exactly 300 words. This is the very first time I gave it a word count and it hit it to the exact number. Now, here's an interesting thing I noticed. I had it just on the regular thinking model, just a standard mode here, and it thought for a minute and 43 seconds to get that right. So, if you have used ChachiPT's thinking mode before or the auto mode, it thought for way less time. So, I'm assuming that before we basically had this light mode every time you used a thinking version. I have never seen it think for two minutes to write anything. Never. Not even close. So, I think the way they made this a whole lot better. It just it really takes his time, which two minutes for this much text is a little bit too long for me, but I like that I have a little bit of control here to pick the different thinking model. Again, this is getting a little more in the confusing side. Again, for regular people using chat GPT, but for heavy users, people that use it like me every single day for the last three years, these are pretty cool options. So far in my first impression, there are clearly big upgrades here, especially if you manually turn on thinking and especially if you extend the thinking time. For the things where I was just testing the regular auto version, they still need work there. And for our Skill Leap members, every time a new model comes out, we go through a huge upgrade process and try to upgrade as many courses as we can, usually starting with our top courses. We just finished upgrading some of our courses with 5. 1 and already 5. 2 came out. So, it's been

### Segment 4 (15:00 - 15:00) [15:00]

really interesting having an education platform where we teach AI with new models rolling out every 3 months. It's been a super challenging time to try to keep our courses updated. But that's the whole value of the course platform is to have something that is up to date as humanly possible. Something the other platforms really just don't do. So, we'll be updating those courses and we have a couple of new ones releasing this month, too. And the last video I posted, I'll put that over here so you could check that out next. I'll see you next