Latest ChatGPT Updates Explained: Memory, o3 & 04-mini, 4.1, Social Media Rumors

Hey, welcome to the Next Wave podcast. I'm Matt Wolf. I'm here with Nathan Lanz. And today we're going to try to do our best to demystify all of these recent launches with OpenAI. Because they've released 4.5 and then 4.1 and then 04 and then O3, and it's just getting really, really confusing. So this episode is designed to break it all down for you, tell you which each of these models are good for, what they're not good for, and we'll even rank dumbest to smartest at the end and tell you which ones to use for which things. So, without further ado, let's just go ahead and jump in and talk about all these recent OpenAI announcements.

Using only 20% of your business data is like dating someone who only texts emojis. First of all, that's annoying, and second, you're missing a lot of context. But that's how most businesses operate today, using only 20% of their data. Unless you have HubSpot, where all the emails, call logs and chat messages turn into insights to grow your business. Because all that data makes all the difference. Learn more@HubSpot.com.

I think probably the best place to start is with this one right here, because this is probably the first piece of news in this sort of sprint that OpenAI is on of releasing news, which is that they just rolled out this memory feature in OpenAI and to me, this is probably one of the coolest, most valuable features they've rolled out. I found it really useful. Yeah, I know. Nathan, before we hit record, you were saying something to the same effect of like, out of everything, the memory is probably the coolest.

Yeah. So it's funny, the team at Grok, they reached out to me about what should they build, like maybe like two months ago, and I was like, focus on memory. Nailed it first. Because OpenAI has like a half baked version of memory. They'll probably eventually get there, but who knows if it's a priority or not. If it's not, you should get there first.

Right, right.

Unfortunately for Elon Musk and Grok, OpenAI definitely got there first. And I think it's going to be one of the biggest things long term. It's the kind of thing where people don't realize why it's important now.

Yeah.

But over time it's going to be clear why it's important. You know, I use chat a lot, like for so many different things, like personal stuff, business, everything. I use it for so many different things. Probably more than the average person. And I asked it to, like, explain, what do you know about me? And I'm not going to show you what it said. Some of it was so revealing, was it not? N. No, it was nice. It was nice, you know, because it knows a lot about my life too. My life is pretty crazy. And just like, some of the insights it drew out of that were just kind of shocking and inspiring. Honestly, it was like, yeah, I have done a lot of great things. I have went through a lot of hard times in my life and all this stuff. And it was just. It seemed to know things about what I want to accomplish in my life, things that I don't think I've told it that were just kind of like, that's me. And so the idea that AIs are already starting to actually understand me, maybe better than I understand myself.

Yeah. And I mean, you've probably had a lot of chats. I know you've mentioned that you've done even like voice chats where you go for like, walks and just sort of like unload your thoughts into the voice chat. And I mean, all of that is in memory now. I mean, and just to be clear, you can turn off the memory. If this feature freaks you out, there is an option to turn it off so like, you don't have to let it remember everything. But like, you know, all of those voice conversations where you're just sort of like brain dumping, that's all in its context now when you chat with it.

Yeah, I think long term that's how AI is going to get really useful for people. Because, like, people think of it right now as just ChatGPT, where you go in there and you copy and paste something in. Maybe have it help you in some way or you just chat with it. But eventually these models are going to understand maybe what you want or what you want to accomplish better than you yourself understand that, or even before you.

Realize you want it or before you do.

Right. And this is like, this used to be like the dream, like when I first got involved in Silicon Valley, like, we were all dreaming of this stuff. We're like, how do we make the web better? Like, well, you got to make it more personalized, actually understands people. It's like, how do you do that? I guess machine learning or something. And like, we didn't know how to actually do it. And people started doing it with like, very basic versions of like, oh, well, Amazon's kind of pulled off something like that because they kind of get what you want. But it's not perfect. Same as Netflix. Not perfect.

Yeah.

Eventually all of this stuff, it's just going to know about the kind of stuff you want to accomplish or do in your life. And I think that's where we're going to get to a point where AI is going to be really helpful.

Yeah. Yeah.

Anything you're wanting to do, maybe it's going to recommend it to you beforehand. Like you said, it's going to know you or it's going to just know the kind of stuff that you want to do and be able to help you better with more personalized help versus just some generic stuff that it says to everyone.

Yeah, I mean, sort of random side topic, but since you mentioned Netflix.

Yeah.

Apparently Netflix is also going to be using OpenAI's AI tech inside there, so it will actually even better understand you and what movies to recommend.

But that makes sense. Like I said, when I first got involved in Silicon Valley, everyone was trying to figure out how do you do that? Because it felt like the natural evolution of software products was to understand users and what they wanted better on a personal level. And no one really nailed it. Like I said, Amazon and Netflix got somewhat close and everyone's just been like tweaking on what Amazon Netflix has been doing for like 10, 20 years now or something like that. So yeah, makes sense.

Yeah. Yeah. Well, I actually like the very first thing when this new memory feature rolled out was there was a button that says do you want to try it? And I said, yeah, sure, let's try it. And it automatically pasted in this prompt, describe me Based on all our chats, make it catchy. This is what it said about me. I don't mind sharing mine. It says you're Matt Wolf, a future forward creator, a high trailblazer, and the go to tech Sherpa for over a half a million subscribers. You're not just riding the wave of innovation, you're helping build the surfboard. Your YouTube channel is a hub for creators, entrepreneurs and curious minds hungry for what's next in AI and tech. You break down the complex spotlight, the useful, and always bring the human angle, all while rocking a retro, futuristic vibe. Whether it's a short, punchy reel or deep dive 20 minute banger, your goal is always the same. Empower your audience to do more, create faster and stay ahead. Off camera. You're a family guy with a white pup named Bowie, a deep love for camping and a soft spot for the San Diego Padres. You've got a musician's soul, a builder's brain and A storyteller's heart. Even when burnout hits or the algorithm plays rough, you keep showing up driven by purpose, not just metrics. You're not just documenting the future. You are part of shaping it. I was pretty flattered by that, so I don't mind sharing it.

Yeah.

Yeah.

I mean, like, parts of that were kind of generic, and then parts of it seemed like it was, you know, kind of personalized view. I have a feeling I probably talked about more personal stuff with it because the stuff it said to me felt more personal than that.

Yeah.

And I also even give it, like, a primer in the custom instructions, like, here's who I am. Here's like, a general.

Yours was probably, like. You keep refusing to get that work checked out. Go get it.

So, you know, I probably provided it a lot more about myself, but it said a lot of things. You know, it's just like reading between the lines. But it said a lot of things that I did not tell it, though, like, exactly like, it. It, like, kind of, like, understood things about me, about how I feel about myself, about my accomplishments, my failures, things I want to accomplish before I die. Like, all these kind of things. It would just kind of shock you.

And it was stuff you didn't actually tell it. It sort of figured it out, some of it.

Some of it. And I was like, holy sh. You know, it's like it's connecting the dots. It's, like, understood my personality somewhat and the things that I get insp. That I want to do or that bothered me or whatever.

Right.

But it definitely connected some dots and said some things that I did not explicitly tell it, which was just, holy crap. That's amazing.

Yeah. Yeah. For me, this one feels, like, pretty on par with, like, the various discussions I've had. Right. I have, you know, Talked about my YouTube channel a lot. Right. I use it a lot for. Hey, I need an idea for this YouTube video. Here's the transcript. Help me come up with a title. All right. Help me come up with some concepts for thumbnails. Right. So I go to it a lot for, like, ideation around my YouTube channel. So it obviously knows a lot about the fact that I do stuff on YouTube.

Yeah.

It has all the transcripts of my YouTube videos because I always copy and paste them in and go, hey, how should I title this thing?

Right.

So a lot of this is based on probably transcripts from my YouTube videos.

Right.

It's kind of random that, like, the one sort of personal thing it picked out about me is that I have A dog named Bowie.

Right.

Not that I'm like, married with two kids or, you know, things like that, but just I have a white dog named Bowie. I'm like, I don't actually remember telling you that, but it's true. So I'm sure it was in like, one of my videos or something somewhere, right?

Yeah. I mean, mine was more personal and, like, to the point where I didn't cry, but it made me feel emotional. Like, reading it, I was like, that is crazy. And just imagine where that goes. Like we always say that's the worst it's ever going to be.

Right, right, right. And the more you use it, the more it's going to know about you and the better response you're going to get on things like this.

We're totally going to have the thing that, like, you know, in the future when you pass away, you'll have like a memory of who it won't be perfect, but, God, it'll probably be like a 90%.

Oh, yeah.

Of the essence of, like, who you were as a person. And that. That is just shocking that, like, that. That technology is like, basically, it's almost there. It's like, it's right there.

It's getting really close. And yeah. You know, this was sort of just the first announcement. This announcement came out on April 10th.

Right.

So this was sort of like the first in a series of like, almost every day.

Yeah. The most important thing GROK could have built, they beat them to it. Small little announcement and then. Yeah. Onto the real things.

Yeah, yeah. They didn't even seem to make much of a big deal about it. I don't think, like, they even did a sort of livestream for this one. Almost everything OpenAI does a live stream about. And this one, they didn't even. They're just like, here, here's a new feature that we rolled out.

Well, they're talking about doing a social network and stuff now, which, if they do that memory is a core feature, that of actually understanding who the people are. Right. That's how you would build some kind of new social network because it'd be highly customized based on the personality of the people in it.

Yeah. I mean, I actually haven't read too much into the whole social network thing. I just pulled up an article from Reuters.

Yeah.

About it. But, yeah, that's another announcement that has sort of bubbled up, is that OpenAI is working on an X, like, social media network, which is interesting. Right. Because there's like a pretty big feud between Sam Altman and Elon Musk.

Yeah, but it makes sense. So this is why I told the GROK guys, like, that's why they should do it. It's because they already have all that data about what people are sharing and how they're interacting with people. And so that would be their advantage over OpenAI. Is that data.

Yeah. So what makes this one different? Have you read into this one at.

All about what OpenAI is going to do?

Yeah. Like, why would people use this over something like X?

It seems like it's all rumors at this point. I'm not sure if it's actually even confirmed. Yeah.

So it looks like it's all sort of speculation at the moment.

Yeah, speculation. I would assume that they wouldn't be going after X, they would be going after Facebook, if I had to guess, because Facebook is that kind of product where it's all about, like, all your friends, all these connections that you have, like, real life. And I feel like OpenAI would be in a great spot to, like, revolutionize that kind of product versus, like, competing with X. Yeah, I'm not sure. I think there's not a whole lot of details yet.

Yeah, super interesting. We'll kind of keep an eye on how that one plays out.

I'm sure Mark Zuckerberg is super happy. Like, his AI effort is totally floundering. And then now Sam Altman may be coming to kill him. I think it's more likely to kill Facebook than X. So, yeah, that's a tough spot to be in.

Yeah. Yeah. So one of the other things that OpenAI has been doing is they've been releasing these new models, but they're also sunsetting a bunch of other models. Right. So apparently GPT4 is going to be phased out on April 30, and then also GPT4.5, which up until this week was like, the newest model they released is also going to get phased out. They sent out an email to all of the OpenAI API users, or maybe it was everybody in ChatGPT, I don't know. But they. They sent out an email basically saying, we're going to phase out 4.5, because 4.1 is better and faster. Like, dude, OpenAI sucks at naming these things.

Yeah. And they had said that they were going to solve that by doing GPT5, which would be like the all in one. The one. You know, the one model to rule them all kind of thing, where, like, you just talk to the one model, and then behind the scenes, it's routing you to whatever is the best model. Probably that's what they're doing, but they didn't do it. They, like, delayed it. They're like, no, sorry, we're gonna have naming problems for a few more months. Like, give us like two or three more months and then, yeah, yeah, that'll be solved. So I kind of Wonder, like, it's 4.5, was it a failed run or are they just finding that it's so expensive that people don't know how to use it, so it makes more sense to improve upon it with some kind of new reasoning model or with the GPT5 to have it part of that kind of stack of things that they might recommend. Like, if you're trying to write something, maybe behind the scenes you're using 4.5, but otherwise it's so costly to use it. Why would you use it?

I don't know. I mean, 4.5 has been my favorite model when I've been using ChatGPT lately, I've been using 4.5 more than any other model. But if I had to guess, I think it's the cost. Right. Because the GPT 4.5, the cost of using it in the API was just, like, mind blowing. Like, nobody was using it because it was just way too expensive. Yeah. So I think cost was probably the biggest factor. And supposedly they rolled out 4.1 inside of the API. So this is all super confusing, right? Because 4.5is available in ChatGPT, but 4.1 is not available in ChatGPT, but it is available in the API.

And they were pushing this, like, good for coding.

Yes, I think they were pushing this one as the best one for coding.

Yeah. And so it's kind of confusing, but, like, then if O3 is out, I assume it's gotta be better at coding. So it's like, what's the use case? Or it's confusing, yeah.

This one is available in the API. This is GPT 4.1 in the API. They released 4.1, 4.1 mini and 4.1 nano. They outperform 4.04.0 mini. Let's see. Coding. It scores 54.6% on Swebench Verified and it's an improvement over 4.0 and over 4.5. So apparently it's a lot better at coding than 4.5. But 4.5 wasn't good at coding in the first place.

Right.

It felt more like human, like when you were chatting with it, I guess.

Yeah.

But it wasn't really a model that was good with, like, logic and coding.

One thing I found interesting, though, is like, yeah, 4.5 has been the best for writing, as far as I can tell. It has the most personality. But they've just kept updating 4.0 behind the scenes. It feels like. It felt like it just kept getting better and better. It's not as good as 4.5 in some ways, but ever since they released even, like the image seg, it feels like they updated the model somehow. Like there's something it's smarter. I'm not entirely sure. You know, they haven't really talked about that. Like, it just appears to be. So I'm not sure what the takeaway is. This is all super confusing. It's OpenAI's fault.

I do have a little bit of a breakdown that we'll get to once we've sort of covered all the announcements that sort of tries to organize it a little bit better. But they did put out this graph, which is intelligence based on multilingual mmlu, massive multitask language understanding. So this one's more of like a general benchmark of just like, intelligence. So we can see that GPT 4.1 nano, the latency, that's the speed. Right. So, like, how fast from the time you give it the prompt to get your response. So GPT 4.1 nano is super fast, but not the smartest. Mini is pretty smart. Smarter and faster than 4. Oh, mini. And the new full GPT 4.1 is about the same speed as 4.0, but slightly smarter. That's what I'm making from that chart.

Oh, my God. That's super confusing. So this is why they need the GPT5. It's interesting. They must have really realized, like, you do need different models for different use cases.

Yeah.

And so they're just like, well, for now, let's just keep giving them different crazy names. But then we'll have to make some kind of smart router. So hopefully that comes soon.

Yeah. Well. And again, I think, like, their ultimate goal is you enter a prompt and then it chooses for you the best model to use. Right.

Like, yeah, well, that's what GPT5, they've already said that's what it's going to be.

Okay. Yeah.

Yeah.

That's kind of the goal that they're trying to get to.

Yeah.

So, you know, so SWE Bench. This is basically how good it is at coding. I believe this is the one where it goes on GitHub and tries to find, like, bugs that need to be fixed and how good is it at fixing them or something like that. Yeah. Given a code repository an issue description must generate a patch to solve the issue. So GPT 4.1 is way up here. But one thing I've noticed IS GPT and OpenAI has completely stopped comparing themselves to external models. Like, you Notice how Gemini 2.5 is not on here? Notice how Claude 3.7 is not on here. We don't know how this compares based on this graph to models that aren't OpenAI's models.

Right? Yeah. And who knows with the coding too. Like, things have changed a lot in the last month. All of a sudden, Google took over coding. Claude was like in a huge lead forever. Then they released 3.7, which seemed good, but they kind of got over ambitious because they would try to do tons of things for you and seem to be making more mistakes than 3.5 did. And then Google came out with 2.5 Pro, their Gemini model. And that's just been amazing. People haven't talked about that enough. Like, Google has taken over coding with AI in the last month. I think most people probably don't realize that, but like that has happened in the last month.

Yeah, no, that's true. And I mean, I still do quite a bit of coding. I've kind of been using Windsurf a little more than anything, a little more than cursor, really. And yeah, I've been pretty much using Gemini 2.5 the most. And every once in a while I'll get hung up and I'll go to Claude and be like, well, can Claude solve it? And sometimes Claude can solve random stuff that Gemini can't.

Yeah.

But I would say like 95% of my coding now is done with Gemini 2.5, which will be interesting because now OpenAI just released O3 and 04 mini, which this new O3 model is their most powerful model. But they also have O4, which is a bigger number than O3, but not as powerful as O3.

Right. And you know, Sam Albin had said a few things about these models. Well, more recently he said the benchmarks you saw for O3, we found some ways to improve upon the model more than we anticipated. So it's actually better than what we showed. But also in the past he hinted at O3 Pro being amazing. And like you said, off camera, maybe O4 is O3 Pro.

Maybe. Yeah, I'm not sure.

Yeah, yeah.

Or maybe O3 Pro might be what they launch as GPT5, who knows?

Yeah, or O3 Pro could be like in a week or two and it's only for the professional, you know, users or something. Like that. That makes sense.

Yeah. This O3 is their current most powerful model that we have access to. It's available to both Pro and Plus users. 04 Mini, I believe, is a faster, less expensive model, but still a very impressive model. Yeah, we could kind of look at some of the charts that they show here. So they actually had another chart in their video that they're not showing here for some reason. But you can see that, like their 04 mini with no tools scored 92.7 on competition math. And they're 03 with no tools scored 88.9. So when it comes to math, I guess 04 mini is better than O3. But then they showed this other chart in their launch video which showed oh3 with python and it got like a 99.5% or something crazy like that on the competition math. Basically, like acing it.

Yeah. That's why they had to have something like GPT5. Like they're going to have to have like one smart router to like. Because all these models are good at different things. And it's, it's, it's. And some of it, you discover it over time. Like, some of it, I'm just like, you learn that like, oh, this model's just slightly better for this one thing.

Yeah, yeah. So we can see here that like O3 and O4 mini, like when it comes to competition code here are like pretty equal. And then down here, the PhD level science questions, O3 is actually better at that than O4 mini, you know, but O4 mini is better at math but not as good at science.

How big is the jump in general from like 01 to 3?

So we've got the estimated inference cost on this axis here. And then this is the math benchmark here. And we can see, oh3 mini, you know, is the lowest cost, but also didn't score as well. Where you've got 04 mini high on math here, pretty expensive comparatively, but also, you know, really smart.

Right.

And then we can sort of see it in comparison here to 03, which is not as good, but it's not showing like the jump from like 01 Pro to 04, which is kind of frustrating.

Yeah, yeah. I think 03 Pro be the more comparable there for 01 Pro. I think they'll probably come out that soon.

But then here we can see when it comes to gpqa, what does that test for? Graduate level Google proof. So basically graduate level questions that you can't just Google and answer for is essentially what that means. So on that benchmark, O3 mini actually performs about the same as 04 mini high, which is interesting. They both perform the same.

Wait, so the GPQA is actually physics, by the way. So it's like physics questions.

Yeah, hard questions.

So cost, performance.01 and. Oh three. Okay, so here we go. We can see. This is math. AIME is math questions. Right. So we can see on math,03 for inference cost is down here. Oh3 high is way up here. And the inference cost is, like, way cheaper than 01 Pro, which I believe is the same as Ohi. Yeah, I don't know why they're using different terminology. I think 01 high is 01 pro.

I mean, I think. You know, remember we talked about when, like, reasoning models first came out? I think when we were out in Boston.

Yeah.

You know, for the HubSpot Inbound conference, you know, we talked about, like, how things were going to speed up now that there was this new paradigm where it wasn't just throwing more data at it, but you also could throw more compute power to. And so we probably would see things improve dramatically faster. Feels like we're now seeing that because instead of the model getting, you know, twice as good in, like, you know, two years, it's, like, now it's like, what, like in three months or something?

Yeah. Oh, yeah. The amount of time between, like, each of these new releases is just, like, compressing.

This is so fast.

But, I mean, here's one that's really impressive. That same GPQA here. Here's the curve of 03. Look at where 01 Pro is basically down here. The cost is, like, the highest possible cost, and it's dumber than 03 Low model, or about on par with the 03 Low model here. And 03 High is still maybe slightly cheaper than 01 Medium. But, like, look at the jump in intelligence way up here. Pretty wild. The speed at which these leaps are happening is pretty crazy.

Yeah. When we first started the show, we were talking about, like, how big a change this was, and I think people probably thought that was hype, and it's actually happening now. And I think people still don't understand it's actually happening and, like, how big of a deal it is. Like, Sam Altman was interviewed at TED the other day.

Oh, yeah, I saw that.

Yeah. He was talking about how I think he said that they have 500 million weekly active users now. And then the TED guy said backstage, you told me that your numbers doubled in the last month since the whole Ghibli thing came out. He was like, that was Supposed to be off record, but yeah, you know, yeah, it's going well. I mean, so possibly OpenAI is what in the ballpark of maybe 700 plus million users. I mean, that is just Silicon Valley has never seen this before. And it's so wild that Sam Allman used to run yc, like the big incubator for Silicon Valley, that he ends up producing the biggest startup of all time. Possibly.

I just wonder why he would want that to be off record. That's something that would probably attract investors pretty quickly to know those numbers.

Yeah, But I think he is in a game with Google and Grok and Anthropic and everyone of like, you want to hype things up, but also you have to keep some cards secret. Right. Like you don't want them to know exactly everything. And there's probably some misdirection and all kinds of stuff going on. Yeah, yeah. It's wild how fast it's improving. And one thing that I found interesting is in the interview he alluded to this being possibly like the most pivotal year for AI, like basically saying. And he's also kind of like saying the best is yet to come kind of thing. And also talking about the next big thing is scientific discoveries.

Yeah.

And the biggest talk on X over the last week that I've seen, a lot of the rumors that were going around and some of them were even shared, like there was like positive things being said by scientists too. Kind of alluding to this being possibly true is that they've had early access to the new OpenAI models and possibly some new discoveries have been made, or at least experiments that they were not able to come up with that are now helping them possibly discover new things in science. So possibly this is like having like really, really material impact on science for the first time ever.

Yeah. No. In fact, let me pull up an article. I was actually reading this earlier. It's on the information here. OpenAI's latest breakthrough AI that comes up with new ideas.

Yeah.

Now AI is getting good at brainstorming. This came out a few days ago before this new, you know, O3 and 04 mini came out. If the upcoming models dubbed O3 and 04 mini perform the way their early testers say they do, the technology might soon come up with novel ideas for AI customers on how to tackle problems such as designing or discovering new types of materials or drugs that could attract Fortune 500 customers, such as oil and gas companies and commercial drug developers in addition to research lab scientists. The apparent improvements highlight the benefit of AI models focused on reasoning, which the ChatGPT maker debuted in September. But yeah, basically saying that these new models, the people that got early access, have been saying, like, this is helping us design and discover new materials and drugs, which is. That's kind of the holy grail of what people want AI to do.

Yes. I mean, it looks like we're already there and we've just like scratched the surface of what this new paradigm with reasoning models can do. But it seems like the next several years it's going to keep accelerating and possibly you know, revolutionizing science. It's like just, you know, it's exciting.

Yeah, yeah. And one of the things about these new models that we haven't even touched on yet is that they're all being built multimodal. Right. So theoretically, you can give them videos, you can give them audio, you can give them text, you can give them all the modalities, images. Yeah, anything. And it will actually be able to understand what's going on within those. But also they all have search now and they're all getting the ability to use tools. And so this chart I found pretty interesting. This is one that they showed in their livestream, but they didn't actually put on their website. You can see that when they gave 04 mini tool use and it was using Python to help with the Math, it scored 99.5 on the competition math. So, like basically acing competition math when you give it access to be able to use tools, it's getting to that point where these benchmarks aren't good enough anymore. We need new benchmarks for can AI find new discoveries for us. How do we make those kinds of benchmarks?

Right. We're getting more and more into a situation too. Like when O1 came out, it took people a while to realize how good it was or what it meant. Sam Altman said this. He said his son is never going to grow up in a time where he's smarter than AI. Like, for a child who's being born right now, who's young, that's just not going to be possible in their lifetime. So, like, that's where we're at. Like these models, not in every single way, not in an entirely general way, every single possible way, but in a lot of ways. They're already more intelligent than most people and that's just going to continue and accelerate. So I'm constantly thinking about, like, what does this mean for me personally in my business? And like, what am I going to, how am I going to use it? I saw Greg Eisenberg tweeting about this other day, if he's telling the Rocky real story, if he was just trying to, like, set up something for social media, but he's talking about how to be kind of in a rut because things are changing so fast and it feels like there's so many different things you could do. And it's like, what do you do in an environment where things are so dynamically changing? We've never experienced this in human history. It's not clear how anyone exactly takes advantage of this, because maybe by the time you do something, it's already changed again.

Yeah, yeah. We're entering completely new territory that nobody's been in yet. And even all the people that are trying to predict when things are going to happen or what's going to happen next are all just making predictions. Like, they don't even necessarily know what's coming or where this is going to lead.

Yeah. I've been preaching this for a while, but I really feel like major corporations are not taking this seriously enough. Being a corporation who has too many meetings and overly, like, moves very slow, that's a bad spot to be in. I think people don't realize how smart Elon Musk has been in this area where he's like, really, like, made things more lean and, like, move fast kind of thing. Every big company should be doing that, because otherwise, how do you survive in an environment where things are dramatically changing?

Yeah. Yeah. Did you see Toby Lutke's statement from Shopify?

Yeah.

Yeah. I mean, he's basically saying what you said, but the big point that he made is if you guys want to hire anybody else in Shopify and you must prove that what you're hiring for can't be done with AI first.

Right.

If you can prove AI cannot do it, then okay, you can make the hire, but you must prove to me that I can't do it first.

Right. And it's also super clever because he's also kind of like, baking in AI into the culture of the company where it's like, look, if you're involved in the company, you have to be thinking about how you using AI to make the company better.

Yeah.

Because otherwise, how could you do that test if you're not, like, actually using AI and like, actively understanding what it could can do? So that's super brilliant.

And I don't think he's coming from a place of, like, AI is going to replace all of you. I think he's coming from a place of, like, this company can be so much more efficient and so much better and get things done so much more smoothly if we're all leveraging AI together.

He tried to make an AI first company, like, even though, even though they're a large org restructuring already to be an AI first company. And most companies are not realizing this. I mean, I think almost all major companies should be reorganizing right now, ahead of time. But, you know, like, people always, you know, they're more reactive than proactive. Like, especially big organizations, like all the companies, they will realize this like five or ten years from now, and then some of them will not survive because of that. But it's gonna be a wild time.

Yeah. So I want to bring this one home with this.

Yeah.

So I actually asked OpenAI's deep research to sort of break down all of the models, like, help get past the confusion, do the deep research on which models are good at what, and then give me like a ranked order of basically which one's the dumbest to the smartest of all of the available mod. Right now, it actually took, let's see, 16 minutes for even ChatGPT to figure out its own dang models with how confusing it all is. But here's the breakdown. So GPT 3.5 is great at basic conversations, but substantially lower reasoning and knowledge than the others. Then GPT4, which came out in March 2023, was a big leap from 3.5. And then you have GPT4 Turbo, GPT4O, which apparently are sort of tied. Then you have GPT 4.1 Nano, which is one that came out this week, which is a very small but surprisingly capable model. 80% of GPT4's ability, basically. Then you have GPT 4.1 mini, GPT4O. So GPT4.1 mini was one that came out this week. GPT 4.0 is the one that's sort of been the default model for a while inside of chat GPT. But GPT4.1 mini, it says, matches or exceeds GPT4O's intelligence. So both are kind of tied. Then you have GPT 4.5, which was the one that they codenamed Orion. GPT 4.5. I know it's been available in the $200 a month plan. I can't remember if it was available in the $20 a month plan, but that one's been one that was sort of more human, like, I guess it's good at writing stories, good at when you just want to have, like a chat with an AI. That was the one that I always fe, like, was the best at Having like a very human, like conversation. Then you have OpenAI's 01. This was their first reasoning model where it actually sort of thought through after you gave it the prompt. And so during inference it sort of did that sort of chain of thought thinking. This was the first model we saw that was doing that. Then you have 04 mini, which was, I believe, one that also came out this week, is a little bit better than just the standard 01. So it's a new efficient reasoning model. When run in high compute mode, it's extremely capable, likely outperforming 01 and approaching the top models in STEM benchmarks. Then you have GPT 4.1 full, which is another one that came out this week, which is OpenAI's latest GPT model, only available in the API right now, not available in ChatGPT. And this one's more tuned for coding and following instructions. But this is why it's a little bit confusing, because we can see GPT 4.1 is actually a smarter model than GPT 4.5. So even though the numbering scheme went backwards, it's actually a more powerful model. And then you have OpenAI 01 Pro, which for the longest time was OpenAI's top dog model. If you got into ChatGPT and you had the $200 a month plan, 01 Pro was the best of the best. It just kind of crushed anything you threw at it. It was really good at coding. Took forever to get into the API, but that was their best for a long time. And then you have OpenAI03, which was also one that was just released this week. It is available in ChatGPT, it's available both in Pro and Plus. So 20 and $200 a month plans. This one is their new best, most powerful model they're making available right now. That's OpenAI03. And then this list says their most intelligent, best possible model you can use is O3 with tool use turned on. So allowing O3 to use tools like Python and things like that to help it with its, you know, math and coding and logic and things like that. When you use O3 with tools, it's their best model. And then also GPT 4.1, they've also sort of tied with it depending on the use case. So insanely confusing. Yeah, but this is the order it ranked it in based on all the deep research.

Yes. I mean, so thankfully, like, people won't have to worry about this soon. So like, hopefully in like 2 months you won't have to worry about this GPT5 hopefully be out. It'll manage all that for you. You just talk to one model. But right now I think their descriptions actually are pretty good. In line with my experience. Like, I think there's like three models people really would. The average people would use right now, like 4.0. That's what most people would use. Like I said, that model keeps getting better.

Yeah. And if you're just using ChatGPT free, that's what you're using. You're using 4.

Oh, yeah, 4o is great. It just kept getting better. They haven't really like made a big fuss about it, but it seemed to have just kept getting smarter in some way. I don't think that's entirely reflected in the benchmarks. 4.5 still seems to be the best for writing. So if you're wanting to edit something or write something, 4.5 is the best.

But that one's also getting phased out. So use it while you can.

Use it while you can. I think it'll still be there in five. I had to guess behind the scenes and then O3, which I haven't used yet, but based on the benchmarks and my understanding how the models work, I would assume that's going to be the best. If you have any complicated questions, like if there's something that you're thinking, like, oh, the model might give me too surface level of an answer, it's not taking the time to think it through. If you're asking something complex, you probably want to hand it to O3 and give it the time to really think through its response to you.

And that's also probably the one I would imagine you're going to go to for coding as well.

Yeah, probably. So, you know, I haven't had a chance to test any of it with coding yet.

There's one last thing that we didn't actually bring up about the new models that they just released, and that's the context window. One of these models is actually a 1 million token context window is a real one.

Like, I remember that was a part of the whole fiasco with Meta as well, is like they said like 2 million context window. And then apparently at all the benchmarks it would like, start to like fall apart after like 128k context or something.

I don't know. I actually kind of trust OpenAI's word on this a little bit more than I trust Meta's word on it.

Yep. Yeah, me too, Me too. Because that was just ridiculous. Like, people tested it. It's like they were hyping up like 2 million contacts or 4 million or whatever it was they hyped up and then, and then people tested it and as soon as it went beyond like 128k context, like the model just like became really dumb.

Yeah, I heard the same thing.

I mean, I'm sure there'll be some diminishing returns like O1 Pro. Even like, you know, that's why I was using a repo prompt is like, yeah, it could take a lot of context, but you do want to kind of be careful about what you give it because if you give it more exactly what it needs, you do tend to get better results back.

But yeah, these context windows are getting to a point where like context length is not going to be an issue for very much longer. Yeah, like you're going to be able to plug in entire code bases and it's going to be able to read the whole dang thing.

Yeah, it is interesting that like Sam Altman kind of admitted, well, he said, number one, like people were, they were asked about like deep seek and he was like, well, has it impacted our growth at all? So not really too worried about it. But also he was saying that like moving forward, it's probably more about the things they build on top of the models. He thinks they're going to continue to be in the lead, he said, but that lead is going to narrow and it's going to be like a, a small lead in the future. So it's all about what they build on top of it. You know, it did seem early on that like open the eye was going to be incredibly ahead of the game, but now it feels like there's gonna be lots of people who are pretty close to them.

So here we go. So as of April 2025, here's the context window sizes for OpenAI's latest models. So it was the GPT 4.1 series, the one that's only available in the API. That one is a 1 million token context window which includes GPT 4.1, 4.1 mini and 4.1 nano. So those are a million tokens. GPT 4O is still 128,000. O3 mini supports up to 200,000 O3 full, likely maintaining or exceeding 200,000. But it doesn't actually say here. 04 mini released along 03 is reduced version of a successor. Specific context window sizes for O4 mini have not been detailed. So yeah, the 03 and 04, they haven't actually said what the context window was. But the new 4.1 which is the one that's going to replace 4,5 is the million token context window. So, yeah, that's pretty much a breakdown of all the OpenAI news that came out over the last few days. There was one smaller update that's just like a quality of life thing. If you are generating Images Inside of OpenAI, they created this library page where you can see all the images that you've generated in just like one spot. You know, kind of cool quality of life thing that they've added as well.

Yeah. Just another move towards killing Mid Journey, possibly.

Yeah, very true.

Yeah. You know, Mid Journey 7 came out like people were expecting it was going to be way better at understanding images. And it's like, slightly better.

Yeah, yeah.

It is still the most beautiful model and it got slightly more beautiful and more consistently beautiful and, you know, and things like that. But it seems like they did not nail the thing where, oh, the characters are going to be completely consistent. Like, you can, like, reuse the characters.

Yeah.

As far as I can tell, they did not nail that.

Yeah. I played around with V7 a little bit. I was actually about to make a video on it and didn't even release the video. Like, I only want to put out videos about stuff that I'm excited about and, like, I was playing with it while recording and, like, it wasn't really exciting me. I'm like, it doesn't feel like that big of a leap to me.

Yeah. The only cool part of it was the turbo thing where you can, like, rapidly try ideas. That's cool.

Yeah.

And you can do voice mode. And so that was actually the only cool thing about version seven was you could just press a voice mode button and just start chatting. And as you're chatting, it starts creating the images super fast. That was cool.

Yeah. I actually didn't play with that mode, so that's one I'll have to mess with a little bit.

That's the actually cool part of it. Like, everything else is like, whatever.

Yeah. I didn't find it that impressive. Ideogram has gotten really, really good. The various Leonardo models have gotten really, really good. Obviously, OpenAI's model is really, really good. Like, all of these models have sort of caught up to each other where, like, midjourney is just sort of in the mix with them as opposed to being the leader, you know?

Yeah, but like I think I said in the past, I think people don't understand, like, with OpenAI's models, with the, you know, with the reasoning models, they're going to be able to actually understand what's going on in the images and the videos they create.

Right.

That's going to be the unlock. Just like making something beautiful. Sure, you can do that, but do you understand what the hell is in the thing you just created? And can you modify that? That's the thing that's interesting. And I think. I think OpenAI is like, the one who's, like, by far the leader in that right now.

Yeah, yeah, they're on the forefront of it, for sure.

Yeah.

But, yeah, I mean, I think that's a pretty good breakdown of everything that's been happening in the world of OpenAI this week. Yeah, we went off on a few little tangents, but, you know, hopefully we brought it all home by sort of showing you the ranking of intelligence. And I think, Nathan, you did a good job of saying, like, all right, these are the three models you're probably going to use.

Yeah.

So, you know, hopefully you found this informative and helpful, trying to declutter your mind on all of these different OpenAI models, because it's still confusing to us. I mean, we're paying attention to it every day and we still get confused. So, you know, hopefully this helped a little bit. And thank you so much for tuning in to this episode. Hopefully you enjoyed it. If you like it, make sure that you subscribe, ideally on YouTube. That's where we're really trying to primarily grow the show. But if you prefer audio, we're available on Spotify and itunes and wherever you listen to podcasts. Thank you once again for tuning in, and hopefully we'll see you in the next one.

Bye, Sa.