GPT 5.4 leaks

There's been tons of rumors and leaks about OpenI's GPT 5.4. And as of today, we can confirm that these are not just rumors. And also in this video, I wanted to take a second to talk about the quit GPT movement. There's a pretty big backlash to some of the actions by OpenAI and and their dealings with the Department of War. Also, quick note, after I recorded today's video, several new things came to light. First and foremost, we already knew this, but the quit GBT movement has been growing rapidly. More people are cancelling their Chad GBT service and switching to Anthropic. Numbers are rather startling. They've basically switched. Entropic is now above OpenAI. OpenAI dropped quite a bit. This is at least if we're looking at estimated firsttime downloads. And as I've posted a few days ago, there's a pretty big backlash against OpenAI. Also, this is hot off the presses anthropic chief back in talks with Pentagon about AI deal. I said there was a chance that this might happen a few videos back. I wasn't sure if that was just me just kind of hoping it would happen, but I did feel like there was a lot of stickiness with Claude and a lot of different departments wanted to keep using it. So, fingers crossed hopefully Anthropic will be allowed back in and hopefully avoid any damaging designations or any sort of repercussions for the government. Here's to hoping that all this turns out well. Unfortunately, also this is again just coming out. Anthropic CEO's candid comments imperil chances of a compromise. Those comments probably refer to this other article that we're looking at. Anthropics CEO's memo attacking opening eyes mendacious Pentagon announcement. Menacious meaning it's bull crap. He's saying their their announcement, the stuff that they're doing is not true. It's meant to deceive quote unquote Twitter morons. A lot of you commented on the previous video. There's tons of discussions. Some of the comments were I don't want to say critical of me, but you guys were speaking your opinion saying that you sort of disagree that maybe I'm being a little bit too neutral. So, I read those comments and I'm preparing a full dive into what we just saw right now because there's a lot that's happening, including on the opening side, there's some sort of a pause while they truly dive deep into potential loopholes and also to construct this safety layer in order to make sure that that this AI doesn't get abused. So, that is to say that in this video, we're strictly talking about GPT 5.4. If you're wondering why I'm not commenting on what's happening right now, it's just because just in the last hour, a a whole bunch of new information dropped that I need to read, understand, and then uh let's discuss it tomorrow. But here we're mainly going to talk about GPT 5.4, which is coming out soon. And it's different in a number of important ways from before from how OpenI was doing this before. the thing that we were all waiting for in GPT5. This GPT 5.4. This might be it. This might be the GPT5 that that we were all waiting for. So, hit thumbs up. Please subscribe and let's go. Yesterday, I think it was OpenAI shipped GPT 5.3 Instant and their next model already leaked not once, not twice, but three times in one week. And it came from OpenI's own GitHub page. code, the error logs, and an employee screenshot. And as of I believe this morning, the information.com, they've they've confronted, they talked to their source over there at OpenAI, and we even have some details about this model, and it looks interesting. So, first and foremost, the rumors were there was going to be a 2 million token context window, which would have been huge. It doesn't seem like that's the case, but it does seem like we will have a 1 million token context window. So let's back up a little bit and talk about exactly what happened. So 5.3 rolls out or 5.3 instant I should say gets updated across chat GBT that's the model that a lot of people use. The instant model is one that sort of a lot of people use because if you're just chatting to the bot you're probably not saying you know think really hard about each thing you're answering you kind of want the fast answer. So now it's the default Chad GBT model for all users. And the big focus is to make Chad GBT less annoying to talk to. So 27% fewer hallucinations. And the other thing that got a lot better is, and before I say it, listen, can you just calm down? Just just calm down. Do do you hate it when people tell you that? Have you ever seen a situation when one person tells another person to calm down and it sort of works as intended where the other person goes, "Oh, you're right, Les. Let me let me just relax." Never has that happened. I feel like calm down is mostly is perceived as an insult almost. It's fighting words. And so apparently the previous models of GPT would tell users quite often to calm down. I actually remember seeing tons of posts on the subreddit chat GPT about this issue. Like for some reason it was really overindexing on that expression like hey calm down buddy. So now apparently GPT 5.3 instant comes with a lot less hey calm down. Like that's been turned down. They got it to uh calm down with using that expression. There's also better writing, better search integration, fewer unnecessary caveats and refusals. I I I I hate those. Those frustrate me to no end. That's a point. I was having a back and forth conversation with one of the Gemini models and something that I asked previously triggered it to really need to have a lot of those caveats and warnings and this weird cautionary language added after everything it said. I might have asked a fitness related question and so thought it had to really be cautious about anything health related. So, so for one of the questions I asked for like, hey, what's a good playlist like a music playlist for cardio? Gave me some ideas and then like three or four bullet points going like make sure you're not playing the music too loud. It's important that you don't disturb the neighbors. If you're wearing headphones, make sure it's turned down to this range of decibb so you don't hurt your eardrums. And like several more like for no reason. decided it really needed to warn me about anything, including listening to music. So, this new roll out of GPT 5.3 instant should be an improvement. That's the official news, but the unofficial news is far far more wild. So, OpenAI engineer Curtis Fjord Hawthorne opened PR number whatever the number was. I'm not going to list it. I'm sure you can find it if you're interested in that sort of stuff, but it was in their public codeex GitHub repo. The issue was that the model, the minimum model to be used was set to 5.4, a model that doesn't officially exist yet. So what happened next was five forced pushes over like 5 hours to change that number from 5.4 to 5.3. So updating the codebase seven times over 5 hours, that's not done to fix a typo, right? So if it's if they just, you know, instead they typed in 5.4 four when they meant to type in 5.3. If that was a typo, you don't update the codebase seven times over five hours for that one typo. That's not a typo fix. It's a fire drill. There's a couple ex users that screenshotted it and next if you want to see it. Another slip came from kind of the same thing. It's a different engineer, a different change. They add a slashfast command to Codex. Codex is open source, so anybody can go on there and see what they're changing in real time. So if there's a slip, people will might catch it. And the original code said toggle fast mode for GPT 5.4. So that's another direct name reference. So that was scrubbed within 3 hours. And yet another OpenAI employee post a screenshot where the model GPT 5.4 is visible in one of the drop downs when you're selecting the model. Post was deleted, but screenshots are forever. Also, journalist Corey Nolles hit some cyberc block where one of the error messages referenced model ID GPT 5.4-B ARM1 10801p codec swip EV3. So that's confirmation that this model is running on OpenAI servers right now. And now Stephanie Palazolo at the information confirms it. So I tend to see that name quite a bit. So this reporter you can say is is bonafideed, right? So well connected has sources that name keeps popping up over and over again. So what do we know about this new model? 5.2 had 400,000 token context window. This new one is confirmed as as far as we know. So the 2 million was rumored 1 million seems to be kind of like the confirmed rumor confirmed leak whatever you call it. I would bet it's going to be 1 million. So that's two and a half times the previous one. So that puts it on par with Google's models. uh and anthropics models who already support 1 million tokens. Another big addition seems to be something that they're referring to as the extreme thinking mode. All right? So you have sort of light reasoning, medium reasoning, high reasoning, extra high reasoning. And apparently if you keep climbing eventually you get to extreme thinking. So this is kind of that inference time compute. It's how long it thinks about the answer before returning the answer. So now we have an extreme reasoning mode. And as far as we can tell, this one isn't just a bump up. This is significantly more time. So, I'd be curious to see what applications there are going to be for this probably research things because it seems like you should expect the answer, you know, in seconds, not even minutes. You should expect the answer like in hours. So, you ask a question and you leave it for hours and you come back, it has the answer to that question. The closest thing that I've seen with chatbots would be like the pro mode or the deep research mode in Gemini where sometimes, you know, it's compiling all the information, doing research. I mean, I've seen it 40 minutes plus, I think, 15 minutes. I personally haven't had one that was running for hours. But also, I don't think that's comparing apples to apples cuz that probably has a bunch of tool claws and stuff like that. This seems to be an extension of just reasoning, more reasoning. So, that's going to be very curious to see. Um, it'll be interesting to see how good that sort of reasoning scaling is working. We just have to figure out what would be good sort of use cases for it, what would be good benchmarks for it. And also this model is better at long running tasks. So tasks that take hours to complete, this model has been improved to be able to run those long horizon tasks. Specifically better at remembering details across many steps, right? So if the if the user has certain details that they need happen or many many different steps, the model can at some point forget what it was doing or forget some minor detail. I love using open claw with for example Opus 4.6 six very powerful model but you run into these issues sometimes they just like face pond like what are you doing when Gemini 3.1 came out I asked to use that model as an API to build something which was going to be just like a benchmark for how well that thing works it came back 30 minutes later saying oh it's done oh but by the way I I I encountered an issue using Gemini 3.1 so I just went ahead and and used Gemini 2.5 cuz it was available and it's kind of like no no no we talked about this is a a benchmark Mark, we're using a specific model to see what it can do. If that model is taking a little bit longer than usual, don't just switch to a different model that defeats the entire purpose of what you're doing. But we probably talked about the purpose of what we're doing at the beginning of the conversation and slowly over time that got lost somehow, the noise or maybe the context window got reset. Whatever the case is, things like that can be very frustrating. You might have seen an AI alignment researcher over there at Meta messing around OpenAI and it just went ahead and just deleted every single email message she had because throughout her working with it, she did early on the conversation said, "Don't delete messages. Just sort of flag them for deletion. Confirm with me before deleting." I think is what she said. And the model goes, "Yes, I'm on it." And just deletes everything. And then she goes, "Why did you do that? I told you to confirm." And the model's like, "Yes, you're right. You did tell me to confirm, but but I didn't confirm. I just deleted them. If I understand the situation correctly, the context window got wiped. So, something got lost that needed to be preserved. So, it seems like from these leaks, these rumors that this new model from OpenAI will be better at not doing that. You'll be less prone to mistakes over long chains of reasoning, which is going to be obviously very, very important for codecs and various coding tasks. So you can think of this as them trying to improve how reliable these agents are over time instead of just raw intelligence. And also you might recall some rumors from OpenAI from a while back months ago I believe about garlic AI. Open eye leader said that this garlic AI or perhaps garlic is a different training process but garlic basically fixed a lot of the training issues that open was having and they said it would help challenge Google's Gemini 3 and and all the Gemini models. So, could garlic be the key, the secret sauce behind these improvements? We don't know. So, it's not confirmed whether or not this is incorporated into this model. But what we pretty sure is going to be incorporated is full resolution images. So, if you're uploading JPEGs, PNG files, WEBP files, what like whatever image files you're uploading before it would it would get compressed. So that means if you're uploading something where the model really needs to see all of the pixels and all the details, sometimes it would do so unreliably because uploading that image would compress the image. So you're you're losing some data there. Here it's going to be full resolution. So if you're taking screenshots of code, if you're doing some sort of medical images, doing diagrams, architectural drawings, schematics, whatever, these new additions, they're probably going to be a pretty big deal. It's also going to have a priority inference system of some sort. So basically, do how fast do you need it? Are you are you willing to use the priority Q? We're not sure, but it's that screenshot that we were talking about earlier that accidentally got committed or whatever that happened that that referenced the slashfast mode. That's what it's referring to. So there's going to be a new service tier system. There's going to be standard and fast. And why this matters, why this is important is if you're using AI agents for some realtime tasks where latency is important, they need to think fast and move fast. Switching to that fast service priority mode is going to be important for those use cases. So this model is running uh internally at OpenAI, right? So we've seen employees having it in their model selector. The information confirms that this model is rolling out soon, so we don't have to wait too long for that. And you might have been noticing how fast they're shipping stuff. At this point, we're seeing a new version roll out pretty much once a month now. And OpenAI is saying they're confirming that this is deliberate. So, this is not an accident. Open AI is making this a priority. Uh, apparently, in order to avoid the hype and let down cycle, so it's something that happened with GPT5. I think we've maybe built it up in our minds to when it actually came out. A lot of people kind of aside as a disappointment, a lot of that might have been just perception, right? How hyped it was versus what actually sort of happened. And that kind of makes sense if you think about, for example, the meter scaling, that exponentially growing chart or they usually show it as a logarithmic chart. But the point is like, you know, there's consistent steady progress. And in fact, it's almost beginning to look like it's it's accelerating. So if all the models and all the points are somewhere along that chart, how could one of them be a letdown? And I guess it makes sense that this could be due to just cadence, right? If you have regular releases, then you sort of expect incremental improvements. There's not too much hype. There's not too much let down. I think they're really trying to avoid the let down. GPT5 was awaited so long that we were just like anticipating it that by the time it came out, I think everybody was expecting something grand. And the reality was again, we were just following that curve. Everything was heading according to plan. The train was right on schedule. It wasn't late. It wasn't early. It was just expectations. So, definitely makes sense that they're kind of doing this, that they're approaching it in this way. It certainly seems like it's going to work. And in terms of the numbers, how well Openi is doing, I'm sure some of you might be very interested in that in light of the quit GPT movement. So, OpenAI is right around 910 million weekly active users. They're hoping to hit 1 billion weekly active users by the end of 2025. So, they're a little bit behind. GPT 5.1 5.2 to updates kind of reinvigorated that growth. Definitely a lot of people went there to use those models because there was a slowdown after GPT5 and now they're taking a hit because of the quit GPT movement or cancel Chad GBPT movement. So with all that said, let me know what you think. Agree, disagree. If you made it this far, thank you so much for sticking with me. My name is Wes Roth. I'll see you in the next