“Anthropic will destroy OpenAI by 2027, here’s why” - Dylan Patel

I think Anthropic, they are a bit of a cult. Anthropic, they left OpenAI for a reason. Their revenue growth rate right now implies that they would beat OpenAI sometime in 2027, which is a really, really big deal in terms of revenue. I think Microsoft will realize they messed up, right? When OpenAI is telling them, hey, we need this much compute in this time frame, the obvious answer from Microsoft, you don't even have $30 billion, let alone 300. How are you going to pay for this? So I think GPT-5 with high reasoning effort is like at least 10 IQ points smarter than Opus 4.1.

A lot of that is because of the reasoning, because GPT-5 high reasons for so long. Opus, it's kind of capped.

Yeah, so the way this deal was done is probably the most unprecedented thing that's ever happened in the history of stocks. It may be the largest contract that was ever announced between two companies. In fact, this is a monumental amount of compute. It is literally many gigawatts. And then in addition, right, not just about securing OpenAI ridiculous amounts of compute, but what was really interesting is Oracle gave a guidance of four years. They told everyone what their revenue is going to be for the next four years. No company has ever done that. A lot of companies only guide a quarter. They don't even guide a year because they're like, 'Ah, not enough visibility.' And Oracle's like, 'No, no, no, no. We see everything for the next four years. This is what we will make.' Which is kind of insane, right? So, of course, Oracle stock popped like 40% and Larry became the richest guy in the world. What's funny is his net worth is a little bit bigger than the size of the OpenAI deal, which is a funny way to look at it. I mean, this sets OpenAI up to have some of the most compute in the world, and they're not tied to the existing hyperscalers, right? Not one that competes with them. Anthropic's with Amazon and Google and those guys try and compete, Amazon poorly. You know, Microsoft's trying to compete with MAI again poorly, but OpenAI sort of has this beneficiary of being completely alone, right? Like, you know, Oracle is not really trying to build AI.

Yeah, I think Microsoft will realize they messed up in terms of like pulling back from the relationship. That's the main thing, right? Is like, you know, Microsoft had the capability to do this entire deal but they were worried, right? When OpenAI telling them, 'Hey, we need this much compute in this time frame,' the obvious answer from Microsoft, right, from anyone who's a finance bro fully and doesn't believe in the technology fully, is like, 'Yo, you don't even have $30 billion, let alone 300. How are you going to pay for this?' By the end of the year, it's like 14 right now, right, billion on an ARR basis, and then you've raised like $30 billion but you've already spent a lot of it, you've already committed a lot of it. How are you going to spend another 30 with me, let alone another 300, right? And that's the size of the deal is 300. So that's the real sort of like why they didn't. Yeah, I think they're starting to regret it.

Yeah, I mean, I think it's pretty obvious to me that OpenAI is the one who's going to be leading the industry. I think there's an optionality for Anthropic, and I think while on the infrastructure side OpenAI is doing more, I think on the model side, growth side, I think Anthropic is potentially ahead and will accelerate past OpenAI, but there's really not room for anyone else.

When you think about Anthropic, right, like their revenue growth rate right now implies that they would beat OpenAI. You know, when you compare both revenue growth rate, and even if you assume both start, you know, slowing down in growth rate, they would still surpass OpenAI sometime in 2027, which is a really, really big deal in terms of revenue. And so, you know, despite OpenAI having a huge head start, Anthropic is right there. That's driven by their models being really good for many use cases, right? So, like, you know, for day-to-day chat usage, for search, for those sorts of queries, ChatGPT is still better. But for the important area of AI for coding, AI for computer use, for productive tasks, Anthropic's ahead. And so that's really the bet they've made. Now, OpenAI is trying to catch up in those areas, but it's pretty far behind still. And as far as like why I think Anthropic, this sort of like bridges into why Anthropic is better in general, right? They are a bit of a cult, right? They believe in the mission much more than any other lab. Not that OpenAI people aren't cultist as well, but when you think about like how focused and locked in they are in terms of AGI, what their vision of AGI is and what they need to do to build it, it's very, very unified. They purely just think, hey, first we have to solve software engineering and then we can use that to build all the other pieces, right? First we have to solve computer use, right, so that it can help us build AGI faster, right? All you need to do is make it so the AI can constantly accelerate your building of the AI, and that's what Anthropic focused on. Whereas OpenAI has a lot of other goals, not everyone in the company's working on that, right? Many people are working on search and many people are working on, you know, consumer monetization and like all these other things. So it's, you know, they just don't have the focus. Why that is, I think, is because Anthropic, you know, sort of has, you know, they had seven co-founders all with like equal equity and these co-founders were like very focused on exactly this mission, right? They left OpenAI for a reason, they were focused, they saw they had a stronger and clearer vision than the leadership of OpenAI. The leadership of OpenAI fully believes in AGI and all of that, right? But the vision on how to get there, why you, what you do to get there is very different and I think that helps. They've got co-founders who are extremely technical whereas OpenAI, right? You know, they don't have necessarily like Sam's not super technical. You know, obviously Greg is a co-founder and he's super technical, right? I completely agree. But like there's also this level of like Greg is much more of a like 'my way or the highway' kind of person from what everyone sort of says whereas the Anthropic, because it's seven people, you know, some of them are, multiple of them are technical, right? There's more dynamicism around that. And so while there's clear vision, there's also different types of technical leadership, right? What Dario is is not what Tom is, is not what, you know, like you kind of go down the list. It's like there are different kinds of leadership and that helps a lot. Whereas like it's like, you know, Sam is who he is and then Greg is who he is, right? And you end up with like the harshness of both has rubbed different people off, right? Some people have been pissed off at Sam and left. Some people have left because of Greg, right? It leads to a different like culture.

Yeah. Yeah. So, OpenAI, that was a big focus for GPT-5 was trying to combat, right? So, that's why when they launched, they had Cursor already, you know, switch their default, right? That was important, right? Because they were focused on code. They were focused on making code better. They're focused on improving the outcomes there. Another one would be sort of like, you know, OpenAI's got Codex. They tried to buy Windsurf. They're obviously like seeing this market. They're just having a hard time truly getting it to, you know, to stick in the same way that Anthropic has, right? And they have all these other focuses.

Absolutely.

Yes, Codex CLI and the extension.

I mean a lot of people say the CLI is much better, right? Because you have Codex the agent in the cloud which is more like for asynchronous work, you can give it 10 different tasks and it just does them in the cloud, but you have the Codex CLI which is more like a Claude Code competitor or the Codex extension which you can use in VS Code or whatever. I think these are much better than actually the cloud because the cloud doesn't really support like a productivity practice. If you're coding, you're working on a single task. You don't want to be like managing 10 different things at once.

But I mean, if I had to compare it like apples to apples, I think GPT-5 with high reasoning effort is like at least 10 IQ points smarter than Opus 4.1.

But I would say a lot of that is because of the reasoning, because GPT-5 high reasons for so long. Opus, it's kind of capped. So I think if Anthropic gave it the same amount of reasoning, I mean obviously it's not that simple but, you know, if they figure it out, I think Opus is much stronger. Like if I want a quick answer actually on the web, I prefer going to Claude and asking Opus because GPT-5 is like you either go instant which you're using a bad model or you go thinking which takes a minute to respond. So like I don't know how we ended up in the situation where ChatGPT doesn't have a single large model that can give you a response within 5 seconds.

Oh yeah, for sure. That's the obvious UI. But I would say like on average, I would say Anthropic models are also more likely to jump to conclusions, right? Like, 'Oh, I see the issue' or 'You're absolutely right,' people are memeing that. Whereas like the Codex with GPT-5 high is way more uncertain, like at least from my testing and I'm using it like four hours a day, is like it's struggling to give you a clear conclusion whereas the Claude models just are way more certain. So it'll be interesting to see like where that goes in terms of software engineering, but honestly I think the move is to use both.

Well, the Claude Code is much easier to use. This is something interesting. OpenAI models, especially GPT-5, cannot really explain things in a simple way. Like this is what I found is that they always use like the most complicated language possible. Claude Code is much easier to work with and I prefer it at like explaining you what the bug is and what's the issue. So that's definitely...

It's happening like day-to-day literally in the last month. Like the Codex CLI has been getting daily updates, sometimes twice a day. You literally have to update it in the terminal. They don't announce it, they don't promote it, and it's been silently getting better. So Codex CLI and the Codex extension in the IDE right now are more powerful than Claude Code. They're harder to use. They're not as polished. You know, Claude Code has been around for longer. So the UI is better. But like if you're looking at a single power, if you have a convoluted deep bug, I definitely would trust GPT-5 high with Codex over Claude Code.

Well, I use both actually and then I just compare them. So like if I have a, you know, I describe what I want to do, either it's adding a new feature or fixing a bug, I have them both analyze the codebase, you know, read all the files and propose a plan and then I give it to the other model, the second model, and I have it rate, right? Like here's the alternative, what do you think of it, how does it compare to your plan. And in 90% of cases they both agree that the Codex one, which is GPT-5 high, is better. So I actually have them, I use them both and they almost always agree.

Yeah, it takes longer. It sometimes reasons for five, seven minutes for sure. So like if you know what to do, like for let's say front end, right? If you want to make quick UI changes, it's better to use Claude Code because you can like boom, 5 seconds, 10 seconds, change that button, change that button. I'm talking about more like, you know, you have a bug, you're stuck.

So, you know, I think the interesting thing is like on this is kind of tells you where their model weaknesses are then, right? Anthropic all around in this range they're way better. OpenAI, you know, GPT-5 is way faster generally. But it's kind of...

And so Anthropic needs to work on both ends and OpenAI needs to work on that middle piece, right? Yeah, and at least Anthropic, right? They said that they're going to release a new Claude pretty soon. And so that should probably help on this front. I think it's like within the next couple weeks, maybe even sooner, but there's a new Claude that's coming out that's supposed to be a good bit better than 4.1.

Yeah, I mean, I have some like info I cannot share but basically they're trying to compete but they don't train their own models. I mean maybe like smaller models, like Cursor trains, you know, the autocomplete models but not like foundational big models. They're more in terms of the integration, right? So like Devin is better if you have a company and you all work in Slack and you can just like delegate small things to Devin and it completes them end to end. So it's more about like the tooling and the utilization of these models which I guess maybe I could flip that question into you is like where do you think will be the gains in terms of AI boosting productivity? Do you think it will be, you know, all the tools around? Do you think it will be the core models getting better? Do you think it will be some of the connectors and environments like MCP maybe? Where do you see the biggest gains in terms of us using it day-to-day?

I think, you know, this is sort of again where you see sort of the difference between OpenAI and Anthropic is tool use. OpenAI is so much better at it than Anthropic, but we're so nascent in tool use. But you could just look at the reasoning traces of OpenAI versus what Anthropic does. But OpenAI will like always call up, you know, you ask it a math problem, it'll or logical puzzle, it'll pull up a Python code itself something and then use that to answer, you know, sort of very different from how Anthropic does it. And so I think like the agency of these things is the real big deal, right? Like how many tools can they use, right? And extending the context, right? Because a lot of this thinking challenge is that context lengths are so low. So how are you going to maintain memory across? So I think that's the big, those two, right? Like across tool use, which tools always ground you back in reality and prevent you from hallucinating so much because they are not a language model, they are like a programmatic thing. So using tools plus extending context and being able to stack the two, and that's the way to train that is really difficult but it's something that they're all working on and there's a lot of like environments that these companies are buying, software development environments, web development environments, all these sorts of things that make it so that they can continue to extend the horizon of tasks. And as that happens, I think that'll really change how people use the models. I think the other aspect of this that Anthropic and OpenAI really need to work on, OpenAI went all in on sort of like reasoning super fast but a lot of their reasoning environments have not been the best, right? Like they focus a lot on competition code and like Codeforces and things like that but that's not necessarily front end. Like GPT-5 is actually bad at front end, right, relative to Anthropic models, right?

And so that's like, why is that? There's got to be some sort of environmental reason why, because they both scrape all the front-end data in the world, right? It's not that, it's something to do with the code environments. And so I think there's a lot of different varied work to do on the environment side for people to actually get, you know, significantly higher model performance, but also that means we're going to see higher divergence in the models, right? Because before everyone used the same web-scale pre-trained data and then they fine-tuned a little bit, right? You know, with like instruction fine-tuning and SFT and all these things. And then as it's gotten more and more to RL, the models are going to diverge more and more and more, right? Not just in behavior, not just in tone, but also like their capabilities, because behavior and tone is things that we've already noticed the difference but the capabilities are what's going to be really noticeable as the different environments that these models are trained in are completely entirely different, right? Like if OpenAI is focused on a bunch of like, you know, competition code and that sort of stuff whereas Anthropic's focused on a lot of like, hey, here's a Salesforce dev environment and you work around with that, like they're going to be completely different capabilities even when you're talking about programming. And I think that's, you know, these are shitty examples that I gave but I think that's the...

Okay. Okay. I think it's a bit of a misnomer though because both are focused on both, right? It's just like, you know, how much focus, how many environments you focus on, how much training time you spend in each.

In those areas there's not much to do in terms of like dramatically improving them. There's rubrics, there's grading, there's all this stuff but like, you know, what models are actually better than most humans at writing, you know, that's a fact. Models are better than most writers, right? That's, I think, that's like actually true. But like the challenge is like to actually do stuff in the writing field, right? Okay, copywriting is fine, but it's about being able to switch between different contexts. It's like the areas that they're deficient, you know, like fine, they can write well, but they can't do all the research. They can't do all the back and forth, right? Like, etc., right? And so I think the main thing is like, hey, if you get them good at code and math, yes, but also computer use, that's the other one.

able to have long context. Then the skill of being able to do good writing, right? In terms of like technical writing, right? Grammatical mistakes, understanding of vocabulary, ability to explain complex topics deeply and concisely, right? These sorts of skill sets will actually be able to be leveraged, which is what the model's already learned on the topic of code, right? Like whatever is happening new, and it'll be able to gather the context it needs to. It'll be able to do the research, right? Deep research is fine. It's really good. I use a lot of deep research, but like there's so much more that can be done if you're able to dramatically extend the context and context switch and use computer use, right? Hey, tell me, can you draft a design document about like why all these things are wrong or what's the problem? Right? And it's like, okay, well, I got to dig through the codebase. I got to dig through the architecture plans. I got to dig through all these things. And that's not something models can do today, right? So that's what I think is an area where combining the ability to write and communicate as well as the ability to code, all these areas are going to start to... right now it's like sort of the developments in coding and developments in writing, right? We got all the writing developments first then and it sort of feels like they've stalled or like are just increasing slowly while the code ones are skyrocketing up. And then computer use and all these things will also skyrocket up. But over time, right now it's like a radar plot, right? Like you're good here, here, here, here, but it's not actually able to bridge the gaps, right? And so that's like sort of specialization versus generalization. So if it generalizes, I think that's the focus. And I think if we have longer context, if we have computer use, if we have very long complex long horizon environments for various tasks, even if they are verifiable, they will start to impart... hey, that's also the same technique you use for non-verifiable domains, right? You go and do some research, you go and grab this, you go and reference these documents. And so even if code is verifiable and medicine is not, to some degree because you now know how to go through and search all these different places, compile information, use all the long context, check your answer, search, etc. And then, hey, you know, at the end of the day, yes, you were able to verify that that was wrong or right, but then you'll also learn to do the same thing in a different field, right? And you'll go through medical journals instead of documentation of code, right? Whatever. But at the same time, it'll generalize to a point where it'll get a lot better at other tasks too, right? Whether it's writing or medicine or whatever it is. Yeah.

Most people suck at prompting actually. I think like actually if you like have you ever like taken someone else's ChatGPT and looked at their prompts.

Yeah, it's tragic.

They use it like Google, you know? They use it...