GPT-5.4 LEAKED: Early Look at the New Features!

So, on March 1st, a few days ago, I posted out a tweet basically confirming that the new model we're going to get from OpenAI was going to be GPT 5. 4, not GPT 5. 3. The reason I found a leak from an OpenAI codeex team member and while he was showing a new feature of the codeex feature, he accidentally leaked the GPT 5. 4 model. Now, something interesting just happened because

yesterday OpenAI made a post and they basically confirmed that the 5. 4 model is coming sooner than we think. Obviously, this was a crazy coincidence or maybe not. Maybe the leak that originally was leaked by the Codex team member could have been a marketing play. But whatever the case is, we do know for sure now that OpenAI is releasing the 5. 4 model sooner than we think. Maybe this week, who knows? If it was this week, I would say maybe by tomorrow, Thursday, if not, maybe early next week.

Now, what's really crazy is that there are other updates that came today and these updates are pointing that the GPT 5. 4 model might be in actually a big jump from OpenAI. So, what are the updates that we might be seeing? First thing, a bigger context window expanded to about 1 million. That means you could submit actual books, big coding projects, and give it access to so much of your content, and it'll make the whole process of processing that content and understanding it much simpler. as well. There's supposed to be a new extreme reasoning mode coming out, which is supposed to have more compute and deeper thinking, which is probably going to be very useful for tasks that involve obviously more thinking, but more logical tasks, more coding projects, things that require the AI model to take some time, understand the problem at hand, and provide you with a answer that doesn't seem rush as well. This new model is supposed to be competitive with Gemini and Claude when it comes to long context models, which is obviously a thing that OpenAI should prioritize given the Claude Opus 4. 6 model is pretty good and it's pretty strong and the new model definitely either has to be on par with it or beat it as well. This new model is supposed to be good at better long horizon task. This is going to be obviously very useful for software developers who might have a project that needs to work in the backend for a couple of hours, might be creating code or whatever the case may be. But this is a feature that's going to make it easily deployable in software development environments. It's also supposed to have improved memory across multi-step workflows. One of the things I hate about AI models at the moment sometimes and I find so annoying is that I have to constantly remind the model of how I work, how my processes are, what kind of tone I want, for example, in my writing, what kind of projects, what kind of things I'm working on. And I want AI models to understand me at a bigger level than it currently does. So having improved memory across multi-step workflows could make that better. But I hope they really solve this memory problem that I'm sure that all of you guys listening agree with as well. There's also supposed to be lower error rates in complex tasks. Now, this is important because AI models are being integrated into really important workflows nowadays. We have them transitioning into enterprise workflows. We have them being used in software development already at scale, but these tasks require lower error rates. This is something hopefully the new model solves. Also, this new model is apparently designed for agent workflows as well automation. Think about codecs, something that they're banking a lot on. They also dropped the codeex model on Windows today as well. So, they really want to go head-on in the agent and automation workflow. So, this new model should solve that problem given that it's supposed to be designed for agents and automation as well. We know that OpenAI also wants to help scientists and research workflows as well. That is why this new model might be good for scientific research and complex problems, which is a good thing. Science-based work, especially given with AI, could help speed up discovery across the world, which is going to be helpful not only for researchers, but humanity as a whole. And this is supposed to be part of OpenAI's shift to monthly model updates. So, we might be getting more frequent updates from OpenAI rather than the ones that we see right now, which is good for us. Maybe that means they are able to take the feedback we provide them, retrain the models, make some upgrades, and push out models that are capable and are meeting the needs of users live, which is going to be very beneficial for all of us and beneficial for me as well because then I have a lot of things to talk about when it comes to these new models. So, I'm happy with that. I hope you guys are happy as well. And now what I'm going to

show you are early tests of the new model. Based on what we know so far, OpenAI has basically publicly made it known that they are dropping a new model. And what it looks like is that people are actually getting AB testing in OpenAI at the moment. And you can see that even this person that released all these leaks that he have, he's saying that if you have a thumbs up or down button next to the coot summary, which basically is saying like what here's my response, opening I will say that do you like this or not? They kind of do AB testing to understand if the new model is meeting the needs of the users which is important before a model release. So we can see the generation from this new model according to this person is a massive jump in coding. It can output about 6,000 lines in one prompt. Obviously the line count doesn't mean anything. At the end of the day it's all about the generations and what the new model is able to produce. So what I'm going to show you are a couple of generations from the new model. We can clearly see that even from the first example, this is a simulation of the Earth and all the solar systems. We've seen this test multiple times and usually it started off with animated looking planets, but now we're actually looking at planets that look very realistic and this was coded using the new model. So definitely this model has upgrades when it comes to coding abilities and generations, 3D visuals and everything like that. Let's take a look at another example. Now, I love doing voxil tests because number one, they're just fun to do. As well, it tests the model's ability to code creative environments. So, if this is what the new model is able to code out, which looks like a Voxil castle and like a town or something like that, I can't wait to test this out myself because I love doing vauil farm style games as well. You know, obviously the basic one Minecraft, which I'll show you as well cuz the new model was tested on that. But this output is definitely the one of the best ones I've seen so far. And if this is one prompt as well, this is kind of crazy, guys, because this new model might be really good at coding as this person says. Then we have the famous SVG PS4 or PS5. Sorry, I think it's PS5. I haven't used mine in a minute, so I kind of forget what the controller looks like, but this SVG test, does the Gemini model kind of do better? Yes. But this is not bad. Like, I'll take this. It looks like the PS5 controller and everything like that. We also see this test being tested across many models nowadays, but it looks like SVG generation, it's a little bit weak compared to what I've seen so far. The newer models, especially the Gemini ones, seem to be very good at SVG generations. And if GPD 5. 4 can create this, it's going to be maybe on par, maybe slightly below when it comes to SVG code generation, but not bad overall. Then we have another procedural city generation and simulation, right? So this is kind of cool. Like we have OpenAI able to generate 3D simulation models which are way better. Like even on this channel I've done a couple of simulations ocean wave one before with the OpenAI model as well. This one looks pretty accurate. Like we have tools that we can add. It looks like we can add like traffic jams recession and things like that. So I guess it's a economy/c city simulator which is kind of cool. But this is once again a really good generation of a simulator. And if this is once again one-shotted crazy, guys. And then the famous Minecraft test. Well, this obviously looks very appealing when I look at it. We don't know how the game play is because a lot of times when I've done Minecraft tests specifically, the game play is where it kind of lacks. The models are good at sometimes generating the visuals, but can I move freely? Is it make sense? Am I able to follow the Minecraft physics, build things, and everything like that? So, we shall only see how good this model is when it comes out. But based on what we have seen so far and based on the claims, this upcoming model might be really good for software developers and for coding tasks and it will obviously translate into other things like workflows. But I hope the new model also does better at writing. What I've seen so far is that OpenAI has become so heavily focused on the coding side of things is forgetting that a lot of people use their models for general purpose tasks as well. And one of the biggest common tasks that people probably use the AI models for are is writing. And what I've seen so far from my personal experience, Claude has done a great job of ensuring that their new models don't sacrifice on writing capabilities. But OpenAI on the other hand, it has lost its edge when it comes to writing capabilities. So I hope this new model not only brings software and coding abilities, but also brings good writing abilities. But that's it for

today's video. Make sure to subscribe to the channel, follow us on Twitter, follow the world of AI, and don't forget to subscribe to our newsletter. We post constantly, and you don't want to miss this. I'll see you guys in the next