AI Image Revolution, Gemini 2.5 Pro & More Use Cases

So this week in Generative AI we had multiple best in class releases, meaning a new tool came out that instantly took the first spot in its respective category. Concretely, I'm talking about OpenAI's image generation, Google's Gemini 2.5 Pro model, and DeepSeek v3. And DeepSeek decided to open source this one again. So just between those stories, there's so much to talk about. But as per usual, there were even more AI releases that you can put to work today. And we'll look at all of them in this Week's episode of AI news. You can use the YouTube show that rounds up all the generative AI releases that you can put to work today. Let's get into it. Okay, so first up we have OpenAI's image generator that you might have heard about at this point. Well, let's just say if you've seen that image in the style of Studio Ghibli this week, chances are it was generated by this new OpenAI update. But as you might know, I already created two videos covering this release. One super short 4 minute 1 and one 20 minute video comparing it to all the competition out there. So if you care to learn more about this release, check out that video. And later in this video we'll be doing a comparison with some of the other releases because this is not the only image generator that came out this week. But for now, I just want to show you one conversation at dawn from our team had with the new OpenAI image generation model. Okay. Created a 3D model of a black Labrador on a transparent background here. Then he prompted for a different view of the Labrador and then he continued to create a screenshot of a video game with the Labrador as a character. Now, as there are two views of it already to be referenced in this, create a character selection screen here and then restyle the game into a 2D pixel art styled adventure game like so. And I wanted to show you this because it really shows off the versatility of this Twitch tool as it's a mix of LLM and image creation model. If you haven't heard about this yet, definitely check out the two other videos on the channel. And we're already working on a dedicated use case video because there's so much you can do in here that I feel like it's worth another video. But now let's turn our attention to the next big release this week and that's the one that was overshadowed by this image generation. It's the Google Gemini 2.5 Pro model, which right now is already accessible through Google's AI studio. As per usual, for a limited time, you can try it out there. You simply switch on over the model to Gemini 2.5 Pro experimental and then you can use it in here, or you can use it in Google's Gemini Advanced, their ChatGPT competitor. But here, you do need a paid plan to access this new model. So what's so special about this? Well, some people consider this the very best thinking model ever released. The only potential competition being 01 Pro by OpenAI. But when it comes to benchmarks, this really crushes everything else on many dimensions. I mean, look at this graph. Compared to some of the other top thinking models out there, like Claude 3.5 Sonnet or O3 Mini High, it's ahead in almost all benchmarks, including an impressive 18.8% on humanity's last exam, the notoriously challenging benchmark that, well, it just wins at. But as you might know, the benchmarks only tell a part of the story. Often it's really about the other things like context, length or the tool use or the vibes of the model and how people adapt to it and if they use it. And this model checks a lot of those other boxes, it's integrated into Gemini Studio. So you get all the tooling that comes with that, which is nowhere close to as deep as what OpenAI has now, but it's solid and maybe more interestingly so, this model comes with a 1 million token context window, which you can see in either the announcement or in Google AI Studio. And more importantly, the performance of this long context window is exceptional. So I want to highlight this benchmark score particularly. It's how these models perform with long context, depending on how many tokens you put in. Okay, so essentially how well they perform depending on the amount of text, because yeah, a lot of models have large context, that doesn't mean they can use it well. And if you look at this line of Gemini 2.5 Pro, you will see that all the way over. At 120,000 tokens of context, it scores a 90.6 out of 100, which is nowhere close to all of the competition. I mean, look at this. Claude 3.7 Sonnet Thinking Anthropics Best Model 53 GPT 01, 53. I suppose GPT 4.5 is better, but that's not a thinking model like this one. I am missing Grok3 in this view and I also would love to see a one Pro on here. So that's really impressive, but at the end of the day, it really matters if people actually adopt this thing, because look, the specs and the benchmarks of this model look really good. But at the end of the day, the OpenAI image announcement completely overshadowed this. By the way, that timing was no coincidence. They announced a spontaneous live stream about one to two hours after Google announced this model. But most people are sort of over this notion of a better LLM. It's kind of hard to deny that a certain plateau has been reached. And even though these benchmarks are better and better for common users, there's no real difference in utilizing these models. Sure, everybody has their preferences, and for coding specifically, I think there's clearly better models and worse models, but if you don't work with code, a lot of this comes down to personal preference. So for example, GPT 4.5 might not have the best benchmarks on everything, but it is my preferred model of choice because I find myself doing a lot of psychological and brainstorming related use cases and I love it for that. The vibes are the best. Okay? And just 24 hours after the release of Gemini 2.5 Pro with all of these impressive benchmarks, OpenAI actually reacted with update to their main model GPT4O. There's a few quality of life increases like fewer emojis, better instruction following, especially with multiple instructions, and improved capabilities when it comes to complex or coding tasks. Something that Gemini 2.5 Pro excels at and also improved intuition and creativity. Now here's the interesting part. As Gemini 2.5 Pro released, that model topped Alamarina leaderboards. But now if this GPT4O update, GPT4O pulled right into second spot again and 4.0 is not even a thinking model. So I don't know what else to tell you except that the AI wars continue here and we're getting better and better products every single week. And figuring out which one of these models will work for you on your specific use cases absolutely goes beyond the scope of this video. I'll be sharing my opinion on these over time as I get more experience with them. Nevertheless, we're talking about this because it's a real advancement, as is the next story, which is very similar by the way, so I'll keep it a bit shorter. China's Deepseek released another model and this is a non thinking model. So this is not the successor to the R1 model. That was the thinking model. This is important. Okay? They released V3. This is their non thinking model. So something that directly competes with, let's say GPT 4.5. And again it crushes the benchmarks. Now the interesting thing about this one in particular is that they released it under MIT license, which means it's open source now. Sure, it's really a give and take on some of these benchmarks if you look a little closer into all of this. But basically it's on par with 4.5 and Sonnet 3.7 and Quinmax, which are three of the models to beat here. But this thing is open source. So yet again, just like with R1, they just put this out there. Anybody who wants can download this model, use it with their own applications. No API, no cost per usage, none of that, just LLM that performs at the intelligence of the best models in the world. But freely available, well, that's pretty wild. And these releases are exactly what quite push the western AI companies to ship like crazy, which they are doing. So let's move on to the next story so you can hear all about this. And the next, let's say cluster of stories comes out of Anthropic. The first one is something that I barely missed last week. It came out first day night for me and I had to record Thursday afternoon. And it's web browsing for Anthropic's Claude. Pretty straightforward as of now, not available in the EU yet. And honestly, even in preparation for this video, I didn't even bother with a VPN or anything because we're going to run our more in depth tests on this thing. But basically every other tool has web browsing now, so they were really just catching up on something that was very obviously missing inside of their Claude chatbot. But more interestingly, I want to cover the release of this think tool that in their words, enables Claude to stop and think in complex tool use situations. Now this is a bit complex, but I can simplify this entire release with the following statement. Right now you have non thinking models and thinking models. The thinking models think before they give you an answer. The non thinking models don't. They just give you an answer right away. Start generating right away. This is a thinking approach that is used with a non thinking model. So the model starts gener output, but once it arrives at something that might warrant some thinking, then it uses this think tool. It sort of just selectively thinks when it's needed. And I think this direction is really the future of many of these products. You're not going to have a model picker that looks like this where you need to know what all of these do best and when to use which one. That's the state of it right now, as they already announced, there's just going to be GPT5 and it's probably going to use something like this in the background where it selectively decides when to think or when to just give you an answer, where to generate an image or when to vibe code an entire application for you. Anthropic has proven to be ahead of the curve on many of these releases. So looking at things that they're releasing like this thinking tool always gives a bit of an insight of where the ball is heading next when it comes to these consumer products. And another exciting development for anthropic is that OpenAI is actually adopting their model context protocol. The shortcut is mcp. You might have heard it before, we talked about it on the show before, across their products. Now this is incredible and something that most people when MCP launched wouldn't consider realistic. Just as a quick refresher, MCP is basically when you give an LLM various tools and it's an open standard. So basically the tools get hosted on a server and then you can connect to that server right here on this machine. I gave it access to a few MCP servers. One of them gives it web search. Another one gives it the ability to actually manipulate files on my computer so I could create folders on the desktop and move files around, things like that. And these MCP servers are sort of standardized tools that will now also be usable with the ChatGPT desktop app and their responses API and the new agents SDK that they talked about. And this is absolutely incredible. This is really a step in the right direction and I was super, super surprised by seeing this because OpenAI has been the most restrictive, trying to maximize all the value they create for their shareholders. So yeah, this is very exciting news. More on this when this actually ships. Admittedly this is just an announcement right now, but it fit in too well with the entire cloud segment. So let's move on to the next piece of AI news that you can actually use. We try to OpenAI audio models that also shipped last Thursday night. So I just briefly want to talk about this, but basically they ship new developer tools to make it really easy to create voice enabled chatbots. And this covers all aspects of voice, text to speech, speech to text and transcription through various models like an updated Whisper and new speech text models including a small one that is really cheap. Basically everybody who has an app and who wants to include voice AI features can now do it really simply at the highest quality level. And I want to quickly share something here because I love to give you a bit of an insight into the different things we do at AI advantage beyond the YouTube channel. One of them is the community, as you know. But the second one is also increasingly consulting with companies who are building products in this space to help them either keep an eye on all the things that are out there, or to bring an informed opinion to the table of what users actually want or which services to use. And one of these that came up over the last week was the various audio models. So we created an entire ranking and it basically breaks down to this comparison table, which I quickly wanted to share with you here. And as you can see, the audio API by OpenAI ranks higher, highest here, next to both Scribe and Sonix. Now, I'm not going to be going into the details here, but basically, if you're a consumer and you want to take advantage of some of these voice features or transcription features at scale, your best bet is probably 11 Labs scribe or this Sonic service, which actually surprised us, but it works surprisingly well. It's just a bit slower than the 11 Labs one. And if you're building app, there's really no competition. To all the new API endpoints that OpenAI opened up with this release. I want to say that with everything that we do at the AI Advantage, which I always pride myself in going a little deeper than we have to, we love to run tests and have multiple people try things out before we give you some sort of opinion. And as the company grows, I hope that the quality of the information that we can provide you through videos like this is only going to keep increasing. But nevertheless, we would be nothing without you, the viewers. So thank you for being here, for being interested and curious. Truly, we would be nothing without you. And let me just speak a heartfelt thank you on behalf of the entire team. I'm pretty confident in saying we all really love doing this and I hope you enjoy watching it too. Now let's move on to the next one. And this is just a quick feature of a use case that I found particularly interesting, and that's Andrej Karpathy, ex head of AI at Tesla. An OpenAI co founder basically built an entire iOS Swift app with AI. And this is something that doesn't get talked about too much. People usually talk about websites, maybe a SaaS business or some WordPress plugin, things like that, but it can also build iOS apps. And he shared the various chats that he had with ChatGPT, where you can see all of his prompting and his project process here, which ultimately results in a legitimate iOS application. Now, he never did this before. So this comes from the perspective of, well, maybe not a development beginner, but a beginner when it comes to coding iOS apps. By the way, if you're not familiar, Andrej also coined the term vibe coding. I think he was the first one who mentioned that and really caught on. And also he's originally from Slovakia, which is actually my home country. My passport is Slovak, my parents are Slovak. I just happen to be very worldly in terms of where I live and how I was raised. But hey, go Slovakia. Actually, fun fact, his second name is sort of almost identical with the Carpathian Mountains, which go all the way into our capital. And the end of that mountain range actually houses the Bratislava Castle on top of his name, sort of. Anyway, I thought it was fascinating to see that you can build a mobile app with vibe coding. And here you have the entire blueprint of how Andre did this step by step. Pretty interesting, if you ask me. Now, on to the next one. If you're viewing this video, chances are you realize how fast AI moves. It's a new feature, a brand new model every single week. And for Mere Mortalist, it's almost impossible, impossible to keep up well. And that's why I want to show you two things. One of them is completely free and one of them is paid. And we at the AI Advantage create both of them. The free one is the LLM rankings that you might already know about, but this really is the answer to the question, hey, if I only have 10 minutes a month and I want to stay on top of AI, what do I do? Well, you check out our rankings that we update every single month. With a quick glance, you'll be able to see what tools are at the top right now for this particular month. And if you're curious about one of them, you can scroll down and look at the reasoning. This is freely accessible and we do this across LLM platforms, image generation tools and video generation tools every single month. Now, if you're looking for more knowledge, that's why we built the community. One of the things we do in the community is we release brand new guides and resources on a weekly basis. You can get a taste for some of these in the free area for Community, where we share various guides and previews of courses we have in there. But let me just go inside the paid area of the community, sort these by the most popular ones, and for example, show you this massive guide on canvas and all the hidden features within there. Now, there really is a lot here, including various little tips and Tricks that you might not have known. And there's a comment section underneath of people encountering and solving various problems or just straight out tips and tricks from the community. Now you might think to yourself, hey, AI move so quickly. What's the point of doing all these guides if they're going to be outdated anyway? Well, actually in March 2025 we went through the work and all 12 members of the team updated every single guide of our community. That's over 150 guides now. Some had to go for complete overhaul, some just needed minor updates. But pretty much two thirds of these needed some editing to them and all of them are updated for today. So whether you want to put yourself into movie scenes or get step by step workflows from the latest features, this is the place to get it. And one final note here is that you can really use these guides in a creative way. And this is something I like to do because for example, here's this big guide on deep research and various use cases that are a bit more detailed than what we do on the YouTube. And then what I do sometimes is I just copy paste the whole guide and then you can go into any conversation here I was creating some banner and you can run a prompt like which One of these OpenAI deep research prompts could I use to improve this? And if I paste the entire guide inside of here, it will look at all of the content and use the step by step tutorials and knowledge inside of the guide. With the current use case that you're working on, you don't even have to read the guide, you just have to know how to copy paste and then it suggests one of the prompts that could work really well with the current project that I'm doing here. So here it wants me to do a competitor analysis to look at some other communities that might have similar marketing materials. What I was creating to here, super smart, no, kind of a smart complementary use case that I might have not thought of myself. And you can do this with every single one of the 150 guides. All of these are handwritten step by step guides with all the details that ChatGPT needs to help you with your tasks. So it's not just that you have access to them by joining the community, it's also that your AI assistant gets access to these and you get all that and so much more inside of the AI Advantage community. That's just a quick little hack that I wanted to share with you today. And now let's get back to the video. All right, next up As I alluded to before, there were actually two more image generator releases besides GPT4.0 image. And all of these are really good at text. So we ran a bunch of comparison prompts, but I really want to take a closer look at one of them, which is this one photo of a highway road with a large billboard by the side of the road that says in this text, written in bold letters on the billboard, cars passing by. Now, the two other models are Ideogram 3 and Reve. I think that's how it's pronounced, Rev. And these are excellent models, especially when it comes to text. So the competition is just unbelievable. So look at them side by side. This is the result from OpenAI. This is the result from Reva Mind the flawless text and the realistic look. And this is the result from Ideogram Text, also looking good. Now, I do have to note that if you run this multiple times, you'll realize that Ideogram sometimes messes up a letter. I mean, it's not a lot, but it definitely happens more often than the other models. 2 out of 4 here were straight up wrong. In ChatGPT, that just doesn't happen at this length. I ran this four times and I got it right all four times. And with Ravi, it's the same deal. The text is just correct every single time. But I would argue that, for example, this image is a bit lower quality. I mean, look at these wooden posts just disappearing into nowhere. And there's a few funky details, like cars on both sides of the road driving the same direction or whatever this is. And you can find these consistently across images. I mean, look at this one, or this car just parked in the grass, or this one with the sign in the middle. Whereas Ideogram looks absolutely fantastic and so does GPT4O image generation. Now, here's a few more comparison images for you between ideogram, chatgpt, gpt4o and reva. So you can make up your own mind. But overall, I would say that just the convenience of GPT4O is hard to compete with if you're already using ChatGPT. The tool clearly is the best at some of these things, and at the very least can hold its own at some of these benchmarks against the competition. Meaning it's equally as good as the rest or better. But it has the ability to edit things with words, and it's right there in ChatGPT. You can blend multiple images together, the entire interface just super intuitive, making it, in my opinion, the best image generator model right now. But hey, the reason these videos exist is so that you can make up your own mind based on the comparisons that we just showed you on screen. All of these are incredible models that you should absolutely consider and just pick the one that works best for whatever you're trying to achieve. And that's everything we got for this week. I'll be hopping on a flight to San Diego in about 10 hours. Actually, I'm going to conference there. And afterwards I'll be going to Japan for two weeks to fulfill the one item on my bucket list when it comes to travel that I never got around to doing, which is experiencing Japan in spring. Now, the only thing that means for these videos is that the background will change, but I'm bringing my mobile studio so you can expect a similar quality and upload schedule as you might be used to from this channel over the past few days. Few years. I haven't missed a weekly YouTube upload in almost four years at this point. Hey, but someone's got to do it. All right, that's it. If you enjoyed it, leave a like, my name is Igor and I'll see you soon.