GPT-5.4 Full Breakdown & AI News You Can Use

Welcome to another week in generative AI. And hold up, actually this week's episode is a little different. So, as you might know, I've been uploading for over five years. Every week, I haven't missed a single one. But a few days ago, I had a surgery. Nothing serious. I just had my tonsils removed. But you're looking at a pre-recorded clip. And what we're doing this week is we're going to still do the news episode. I'm going to still work on it, but because I'm not allowed to speak, we're going to have AIore, this guy. Hello there. Present the episode to you. The writing still comes from me. I just wanted to let you know that this is a onetime thing. And I hope you enjoy this. AI powered experiment. Let's begin. So, here's the

deal. Last week, we gave you a really quick update right before the video went live because OpenAI dropped GPT5. 4 right at the end of the week after we'd already wrapped up the production on our Friday video. But now, we've had a full week to test it out. And today, we're looking at what the community is doing with it and then showing you the results from some of our own benchmarks to see how it stacks up against Gemini 3. 1 Pro and Claude Opus 4. 6. Let's start by looking at what people are already building. First up, OpenAI launched a developer showcase site that shows off some stuff people have created with codecs running GPT5. 4 and it's interactive. You can actually click on examples to test them out. There's the Rift Vox, which is an entire firsterson shooter based right in your browser. It's no Call of Duty, but it's an impressive example nonetheless. Next up, we've got some impressive SVG animations from Peter Gstiff and Dev over on X. Peters is maybe the most visually impressive animation I've seen any AI create yet, but I really like Devad's post because he's got the comparison to Opus 4. 6 right there. So, you can actually study the differences between how the models handle it. Next up, we have this theme park simulation game made by Chris. It's basically one of those clicker games, but it also reminds me a bit of Roller Coaster Tycoon with the ability to actually place the buildings wherever you want and see how it affects the flow of traffic. And finally, a really funny one. A user named clad3815 asked GPT5. 4 to draw the OpenAI logo in Microsoft Paint to test its computer usabilities. The first drawing was awful and GPT actually recognized that. So it opened the browser, went to Bing images, found an OpenAI logo and then used the Windows screenshot shortcut to just snip the logo and paste it into Paint. So it completely cheated, but ultimately it did get the job done. Now all of that is impressive, but the best way to figure out how good these tools are is to put them to the test. So that's what we did. Let's start with our design test where we asked the top models to create a visually stunning design website for a studio that would impress web front-end developers. In practice, Gemini 3. 1 Pro and Opus 4. 6 kind of tied for first place here. GPT 5. 4 was clearly a step behind them on this one. Next up, SVG generation, a Death Star over LA. Claude won this by a mile, nailing the entire scene. GPT 5. 4 and Gemini both completely messed up the Death Star with Gemini also missing the city lights. Then we tried creative writing. In practice, Gemini was straight up bad and boring. GPT 5. 4 and Opus 4. 6 both wrote genuinely interesting stories. Obviously, this is subjective, but I enjoyed GPT 5. 4s more, so it takes the win. Now, let's talk about research. We asked for intensive research on the current state of copyright law regarding works created with AI tools, explicitly asking for a massive report, and GPT 5. 4 took this seriously. It spent 4 minutes just thinking and then 4 minutes writing. It checked sources across the internet for worldwide context and delivered exactly the massive report we asked for. Opus 4. 6 didn't spend quite as long thinking, but it still delivered a comprehensive report and came to basically the same conclusions. Gemini unfortunately spent the least amount of time thinking and gave the shortest report, which I consider a failure in prompt adherence since we specifically asked for something massive. And finally, coding a game. We asked them for a 3D synth wave spaceship game in a single prompt. Impressively, all three actually made playable games, but Claude is the clear winner here by a mile, delivering a functioning game with obstacles and a score. GPT 5. 4 and Gemini kind of tied for second. GPT 5. 4 had more detail, but messed up the ship's orientation, and Gemini was just too basic. The bottom line is this. GPT 5. 4 is incredibly capable, especially when it comes to deep research, reasoning, and creative writing. Claude Opus 4. 6 6 is holding its own and completely dominating when it comes to specific coding tasks and SVG generation. Gemini 3. 1 Pro had some good moments, especially in design, but it definitely struggled to keep up in the heavier text and logic benchmarks. If you have not tried running your own sideby-side tests yet, I strongly recommend it. It's really worth exploring to figure out which one fits your specific workflow best. All right, now let's see what's next. Hey there, I

would love to invite you to an event that I'm hosting in March. IBM invited me to teach in an upcoming AI webinar. So, if you're interested in AI and don't consider yourself an advanced user, then this is designed exactly for you. Heck, even people who are deeper into it can benefit from a refresher on context engineering. Because the title of the session is context engineering for students, making IBM Granite your academic co-pilot, and it will be a tight 60-minute session teaching you the foundational concepts around context engineering, and that is the most essential skill when it comes to working with AI. Understanding context is absolutely critical with modern AI tools. and I'll be sharing what I learned over the years. In the webinar, we'll also be showing off some IBM tools like Watson X and Granite. Plus, I'll be sharing a bunch of free resources from IBM's skill build. So, if you're a student looking to get the most out of AI, you should totally come check it out. But also, if you're not a student, maybe you're 30 years into your career or you're a student of a new topic, you're also more than welcome to come. It's going to benefit anybody who's looking to upskill and wants to use AI on that journey, which let's be honest, in this day and age, how do you not use AI on a learning journey. So, come join me on March 26th, 2026 at 5:30 Central European time, which is 11:30 a. m. Eastern time. And the link is in the description below. Hope to see you all

there. Next up, we've got a new feature on Canva called Magic Layers that lets you turn any image into easily editable layers. This is an incredibly powerful tool if you're doing any kind of design work like creating social media posts or YouTube thumbnails. And since Canva is primarily used by designers, it's a perfect fit for the platform. The technology itself isn't new. In December, we covered a tool on this show called Quen image layered that does the exact same thing, but we haven't seen this tech integrated into any popular platforms until now. And once you see it in action and start using it, it starts to feel like one of those things we'll look back on in 10 years and wonder, how did we ever do design work without this tool? I do want to mention that Magic Layers is incredibly good with infographics and digital design in general, but it struggles when you ask it to work with realistic images. So, make sure you understand what you should and shouldn't be using this tool for and then go give it a shot. You can test Magic Layers for free on Canva, but if you want to use it long-term, you'll have to sign up for at least their bottom tier paid plan, which isn't too bad at only $15 per month. So, go check out Canva Magic Layers if you're interested. And now, let's see what's next. Okay, so let's move on to

Microsoft, who just introduced C-Pilot Co-work, which is a new Microsoft 365 feature built directly on Anthropic's Claude system. If you're not familiar, cloud code was very cool, but far too technical and complicated for most people. Then Anthropic made claude co-work as the next step and it was more useful for more people but still fairly limited to your desktop. The idea here is that this release from Microsoft is the next step in that process. You get the exact same capabilities as claude co-work but wrapped in Microsoft 365's enterprise security and it isn't limited to your local files. It operates in the cloud and runs tasks in the background pulling from your emails, meetings, files, and chats. You just describe an outcome and co-work breaks it into steps to produce actual deliverables across different apps like slide decks, briefing docs, and workbooks. Right now, this is available in a limited research preview and it comes tied to a new $99 per user enterprise bundle. But since many businesses worldwide use Microsoft 365, a lot of people are using Copilot, even if it isn't one of the best AI systems, and this is going to be a massive upgrade for those people. So, if you're using Copilot for work already or you're interested in adding this kind of system to your business, go check it out. Hey, if you're enjoying these stories, make sure to subscribe. It really helps the channel and it only takes a second. Okay, now let's continue. All right, so next up, we've

got two updates to Google's Notebook LM. This first one is available for everybody, even people on a free plan, and it's an upgrade to the existing infographic function that they added last year. Now, in addition to creating an infographic based on your notebook sources, you can change the style of the infographics. You just open up infographic options. Choose a visual style like sketch note and hit generate. If you don't like one of the preset options, you can create your own by just typing out a description of what you want to see and it will make that too. So, not a massive update, but Notebook LM is continuing to grow in popularity. And they're adding upgrades like this all the time. So, this is your reminder to go check Notebook LM out if you haven't already. Next, we've got another Notebook LM update with the launch of cinematic video overviews, which basically turn your source material into more polished explainer style videos. And I really like what Google built here. It doesn't rely on pre-made templates, but instead it approaches visual creation more like a director with the system analyzing and evaluating the material to decide on the most appropriate narrative structure. Then, it decides which models like Nano Banana 2 or VO3 should be used for each part of the task. As you can see from these examples from Ethan Mllik on X, it creates some beautiful videos all based on the source info he provided. Robert Scobble also posted an example and kind of said it best with how he closed the post. Do you all get how insane this is for learning about new things? He's right, but as of right now, it's only available for those on the expensive Google AI Ultra plan, which runs you $250 a month. So, if you really want to test this out, feel free to drop the money on it. But my honest recommendation would be to wait to see if they make this available on the cheaper Google AI Pro plan at $20 a month. All right, on to the next one. This one was slightly disappointing.

Luma just released Uni1, which is their first model that combines reasoning and image generation into a single architecture. Luma has been doing awesome work in the multimedia AI space over the past 3 years, but this particular release doesn't quite hit the mark for me. They claim that benchmarks show it performing better than Nano Banana 2. But if you take a look at the examples Luma put directly on the Uni1 website, it's clear that it just isn't on the level of Nano Banana 2. Companies always cherrypick their absolute best examples when they release something. And if these examples aren't impressive, the standard outputs are definitely not going to be impressive either. So at this point, we can't recommend Uni1, but it's an interesting concept, and it doesn't mean that Uni 2 or Uni3 won't be amazing. Every system starts out as the worst version of itself, and Luma has proven they know what they're doing in AI. We will keep an eye on how this develops. Okay, so I want to do a quick

follow-up on the Anthropic versus OpenAI situation. Long story short, the US Department of War wanted to use anthropics technology and Anthropic agreed as long as the Pentagon respected two red lines, no autonomous kill systems and no mass surveillance of the American public. The deal fell apart as the US Department of War wanted no restrictions on the use of anthropics tech. Days later, OpenAI stepped in and agreed to their own deal with the Department of War. And this deal is turning into a real mess for OpenAI. The day after the deal was announced, uninstalls of CGPT's mobile app jumped 295% in the US. At the exact same time, US downloads of Anthropics Claude app jumped 51%. A lot of people are unhappy with OpenAI's decision, and if you're one of the people looking for a chat GPT alternative, I strongly recommend checking out the standalone video we made on how to switch from chat GPT to Claude. Next up is another quick

follow-up to last week's story where we talked about kids using AI to cheat on homework. So, OpenAI just published a study in collaboration with Stanford University and Estonia's University of Tartu that is designed to measure how chat GPT impacts student learning and knowledge retention over time. They ran a trial with over 300 students and found that microeconomic students who use Chat GPT study mode actually scored about 15% higher on their exams compared to those who didn't. To be fair, the results in other subjects weren't quite as statistically significant yet, but in practice, it shows a lot of promise. I just wanted to share this because this data shows that if students use these tools the right way, it is arguably more effective than the ways we used to study. Just another piece of evidence to show that outlawing AI tools for kids isn't the answer. Teaching them how to use them properly is. All right. Next

we have Netflix buying a stealth AI filmmaking company called Interpositive that Ben Affleck actually started back in 2022. We've been hearing for years that AI is going to change film making forever. And a lot of people think that means we're going to be feeding two sentence prompts into a computer and getting entire movies out of it that go directly to theaters, but it will be more along the lines of what Interpositive does, handling really tedious post-production stuff like relighting, swapping backgrounds, removing items to fix continuity errors. So anyway, I just wanted to share this one as it's a good counterbalance to the narrative that AI will replace artistic talent in Hollywood. At least for the near future, it's going to look a lot more like this. creative talent doing what they do best and new AI tools speeding up boring background processes.

And finally, we have a really practical study from Anthropic about how AI is actually impacting the labor market. Long story short, they built this early warning system to see what jobs are genuinely getting automated rather than just guessing. If you're worried about how AI might impact your job prospects, you should go check this study out. And here is what I recommend. Download the entire PDF, toss it into your chatbot of choice, tell it your exact profession, and ask it what you should be doing to prepare for how AI is changing things.

And that's pretty much everything we have for this week's episode of News You Can Use. I would love to hear which one of these was your favorite. And other than that, I hope you have a wonderful rest of the week. I'm AI Eigor, and the real Eigor will see you very soon.