What's going on with Cursor? (The OpenAI disrespect is insane)

0:00 OpenAI posted a blog that at first glance might not seem like a big deal. Our decision on Cursor following its acquisition by SpaceX, but the drama that preceded it, the skirmish between

0:14 XAI employees and OpenAI employees and what this means for the future of AI makes this story far more interesting than what you might think at first glance. Okay, so the first thing you

0:26 need to know about this blog post is OpenAI dropped it at 700 p.m. Friday night. That should clue you in as to exactly what they were trying to do. Whenever a company makes an announcement

0:38 late on Friday night, you know, they don't want people talking about it. It's going into the weekend. People aren't watching news as much as they usually do. They're not browsing social media as

0:48 much as they usually do. So, they really tried to get this out at a time where the fewest possible people would see it. But, of course, I saw it immediately and I literally gasped. This is a big deal.

1:01 So, let's get into it. What does this mean? Today, we notified SpaceX that we intend to wind down our contract providing open AI models to Cursor. Yes, Open AAI is no longer going to be

1:15 available in Cursor. Their partnership of which cursor pays OpenAI a ton of money to provide their models in the Cursor platform is now over and there is well over a decade of history that led

1:32 into this decision. This is not an easy decision for any company to make. OpenAI is giving up a substantial amount of revenue. This potentially looks pretty negative to anybody looking on to it. It

1:44 looks spiteful to say the least, but I'm going to break it all down for you. This decision was incredibly tough as we care deeply about our models being broadly available for developers. We are making

1:55 this choice and just brace yourself for this because we cannot be confident that SpaceX will use our technology within our terms of service. Ready? based on our experience with Elon Musk's

2:10 companies violating contracts. They literally called out Elon Musk breaking previous contracts and a history of breaking the terms of service as reason for pulling their models from the cursor

2:22 platform. But why are they nervous about that? And what did Elon do to make them so nervous? So to understand how we got here, you need to understand a little bit about the Elon Musk and Sam Alman

2:34 beef. So I'm going to give a very brief overview of the history here. So in 2015, OpenAI was started. Sam Alman was the main guy there. Elon Musk gave them donated to them a ton of money. And it

2:49 was a donation because OpenAI was a nonprofit. Then in 2017, OpenAI figures out, oh, we're going to need a ton of compute to compete with Google DeepMind. And that was their main competition at

3:01 the time. And so how do we get a lot of compute? Well, we need a lot of capital, which means we need to be able to raise money. And so then around 2017 2018, there was a falling out. Elon Musk

3:13 wanted to take over OpenAI as CEO. Sam Alman basically said, "No thank you." And Elon Musk left. In 2019, OpenAI starts operating under a cappit structure, basically allowing them to

3:26 make money and to potentially pay early investors for their investment. Then of course we all know 2022 2023 Chat GPT explodes out of the gate. One of the fastest growing products of all time

3:41 changes the world. Then later in 2023 Elon Musk launches XAI, a direct competitor to Open AI. And then in 2024, Musk sues Open AAI saying, "Hey, you can't do that. You can't take a

3:54 nonprofit, take my money, and then suddenly convert to a for-profit and become one of the most valuable companies in the world." They start going back and forth with all of these

4:03 private emails that now become public. They go to court. It's ultimately dismissed, but there's a lot of spite. And this is a big deal because having access to multiple models and having

4:15 multiple different models working together is incredibly strong, especially in an adversarial way, actually produces much better results. And the sponsor of today's video does

4:26 exactly that with their code review project. Now, GPile is used by some of the biggest companies in the world, including Nvidia, including Meta to review the code before it gets on

4:37 production. And so, this is GPile T-Rex. And so, you can see real world examples from these big companies, open source projects. I'll link it down below, but check this out. So, this is for the

4:48 Axios project. It actually found a security loophole and called it out. This is a P1 loophole. It gives you a prompt which you can just copy and paste right into your agent to fix it. Look at

5:00 that. One of the maintainers valid and fixed. So, Greile caught something that would have otherwise made it to production. I'm such a big Reptile fan. They've been fantastic partners. I use

5:11 Greile on every single one of my projects because there's just no way to review all of this AI code without some additional help. And Graptile is the best way to do that. So, I'm going to

5:22 drop a link down below. Go check them out. Thanks to Grappile for sponsoring this video. Let's get back to it. Okay, so that's the very brief history. Now, going back to this blog post. Now, they

5:31 point out two times where Elon has broken contracts, specifically after Musk acquired Twitter, now part of SpaceX. The company broke the terms of our contract alongside many others.

5:46 Under oath earlier this year, Musk admitted that XAI, now also part of SpaceX, had violated OpenAI's terms of service, specifically around distillation. And if you don't know what

5:58 distillation is, it's actually quite simple. It's basically the process of asking a large AI model a bunch of questions, getting the answers, and then using those questionans answer pairs to

6:10 train a smaller model. And what that allows a company to do is basically leapfrog ahead, skipping all of the tedious work of collecting training data, curating it, cleaning it up,

6:22 making sure that it's really highquality data, and just going straight to this quote unquote teacher model to teach the student. And then there's this thing called distillation attacks. You've

6:33 probably heard of it. Anthropic posts about it earlier this year. They basically claimed China and the Chinese AI labs were doing industrialcale distillation attacks on their models.

6:45 The Chinese labs were basically spinning up hundreds of thousands of accounts, asking a bunch of questions to the clawed models, getting the answers and training really good almost as capable

6:56 models for a fraction of the price that it would have cost to do it organically. And so what OpenAI is basically saying, "Hey, now that SpaceX aka Elon has acquired Cursor, if we let Cursor use

7:10 our models, they're capturing all of our data." And they might go and train a future model using our data. Now, they don't want that. They've poured billions and billions of dollars into training

7:24 GPT 5.6 and Astra when it eventually comes out. And by allowing cursor to have all of that data, they are saying against their terms of service, cursor, the SpaceX team, etc. could train a very

7:40 capable, nearly as good, if not as good model for a fraction of the price that it costs OpenAI. And that's anti-competitive in their mind. and again more importantly against their

7:50 terms of service. And they specifically said, "Well, we're not worried about that for other companies, but Elon, Elon specifically has this pattern of breaking contracts." All right, so let

8:03 me show you now some of the skirmishes that happened on Twitter over the weekend after this announcement was made between XAI employees and of course OpenAI employees. So first, this is

8:15 Michael Truel. He is the CEO and co-founder of Cursor. By the way, I should say right now, I am a big fan of Cursor. I'm also a very big fan of OpenAI. I do have relationships with

8:30 both companies. I have existing business relationships. So, I'm going to try to present all of this news in the least biased way possible, but I like both of these companies. So, hopefully that

8:40 cancels it all out. So, Michael Truel from now XAI cursor. We're sorry to see that OpenAI put out a note saying they plan to block cursor users from accessing OpenAI models in 3 months. And

8:52 if you're not familiar with Cursor, it's an AI coding platform, but historically it is only the platform. They haven't built their own models. So they were very dependent on Anthropic and Google

9:04 and OpenAI to actually power the platform. Now they have had their own second tier model for about a year now. And also now that they're acquired by XAI, they have Grock, but really when

9:17 you come into Cursor, you're looking at cloud models, you're looking at OpenAI models, and that's what's powering it. All right, so here's now where a little spice comes in. Michael Truel, Open AAI

9:29 models serve about 5% of cursor user traffic. Only 5%, which is crazy to think about. I would have thought it's a much bigger percentage. But you have to think what is he talking about? Is he

9:42 talking about total token usage? Is he talking about total spend? Cuz those are very different things. The total token usage is probably much more skewed to the smaller, higher efficiency models,

9:55 maybe open-source models, but the revenue is certainly going to one of two companies. It's either Anthropic or Open AAI. they charge a lot more for their tokens and their tokens are typically

10:08 better. However, he does say we're speaking with the OpenAI team to resolve this. Although I don't think there's going to be a resolution here. The 5% number, like I said, is kind of a weird

10:18 thing to say because it's probably token usage, token volume, not necessarily revenue. And in fact, Tibo said it right here. Tokens are not a proxy for revenue, nor value created. in the open

10:32 AAI models are on the very frontier of token efficiency. Smaller or less strong models require many more tokens to achieve a task and therefore will inflate traffic share significantly. So

10:43 he's saying, "Hey, our 5% is probably worth more or at least as much as all of your other smaller, less capable models, including open- source models, which it's kind of hard to disagree with. A

10:55 lot of work can be done with those less capable models, but ultimately you do need to pay for the best of the best at a certain point and you will be paying. They are not cheap. And so, one more

11:06 point on why they're doing this, why OpenAI is banning Cursor from using their models. A few months ago, it was announced that SpaceX SpaceX AI is acquiring Cursor. And this, I've said

11:21 before, is probably one of the best acquisitions of all time, one of the most strategic, one of the most synergistic. And there's a reason. Cursor had insanely valuable data.

11:34 Basically, any time that you have built anything using cursor, they have all of that data, incredible, hugely valuable coding data that they can use to train future models. And then

11:48 you had XAI or Space XAI which basically came out of the gate with a really strong coding model but really fell short in the long run. Their coding models were not that good. It wasn't

12:01 really being used. But what they did have was insane amounts of compute. Space XAI invested heavily into buying Nvidia GPUs. Initially 200,000 building out their Colossus data center, then a

12:17 million building out Colossus 2, and they're going to keep going. And they've built more compute capacity more quickly than basically any other company on Earth, which is a crazy feat in itself.

12:30 But what happens when you have all of that compute and a model that the market is not really that excited about? Well, you have a bunch of idle GPUs. they didn't have enough demand to match their

12:44 GPU or compute supply. So, they had all of these idle GPUs. Then, they look over at cursor and they're thinking, "Wow, you have a ton of data. Plus, you've obviously proven yourself to be able to

12:54 build models cuz, you know, they had a pretty good coding model. It wasn't the frontier, but it was a workhorse model." And uh, hey, maybe we should get together. Maybe we should do this thing

13:04 together. you come on over here, bring all that data, much of which came from Anthropic and Open AAI and let's build some models together. You can use our GPUs. And so they completed the

13:16 acquisition and now they are building incredible models. Grock 4.6 is a very, very good coding model. Now it's not the absolute frontier like a Fable or GPT 5.6, but it's really good. And I bet

13:30 Grock 4.7, which is not out yet, is going to be incredible. I bet it's going to be nearly at the frontier, if not matching the current frontier. I can't say anything about the next version of

13:42 Fable or whenever GPT6 comes out, but it's going to be really good. So, that brings us back to this big announcement that OpenAI tried to sweep under the carpet late on Friday evening. Open AAI

13:54 is banning Cursor from using their models directly. Now, you can still take your OpenAI API key, load it into Cursor, that's fine. But basically, this kind of native support that Cursor has

14:07 offered for basically every model provider is no longer allowed. So, this is unprecedented, right? No, it's not. And that's what makes it even crazier. There are so many contradictions

14:19 happening right in this moment. Last year, Google bought Windsurf. Windsurf was basically a lot like Cursor. It wasn't nearly as well-known. It wasn't nearly as big as Cursor, but in a lot of

14:33 ways, it sat kind of in the same place in the market. They didn't really make their own models. They were dependent on other model providers to power it. And they were AI coding platforms, but just

14:45 that. And then Google bought Windsurf. And then June 3rd, it's reported in TechCrunch that Anthropic is limiting its direct access to claim models. So basically, Anthropic was like, "Oh,

14:59 Google just bought Windsurf. Okay, well we're not going to give you our data, so now you can't use it." Sounds very familiar, right? Also, last year, Anthropic revokes Open AAI's access to

15:10 Claude because Anthropic was thinking, "Hey, we're really worried about distillation attacks, and if Open AI is using our models at OpenAI, why else would they be doing it other than trying

15:23 to get our data?" Well, it's called benchmarking. And it's basically saying, hey, I have a really good model. I want to see how good your models are. Let me just test it against the same benchmark,

15:32 the same test. But they thought, okay, I don't want to give up my data. And in fact, we have better coding models. I don't want you, my competitor, to use my coding models to catch up to me. Okay,

15:43 but that's not it. In fact, in January 2026, it was reported that XAI had been cut off from using anthropic models. So Anthropic back eight months ago said, "Hey, XAI employees, same thing. We

15:59 don't want you using our models. We don't want you to have a better coding model to allow you to build infrastructure to catch up with us. We definitely don't want you distillation

16:07 attacking us." So this was already a thing. So Anthropic back then banned XAI from using their models. Okay, now fast forward. Of course, they would probably do the same thing today, right? I mean,

16:20 Anthropic banned XAI from using their models back in January, so why wouldn't they continue the Oh, whoops. Nope. It's all good. Tom Brown, co-founder of Anthropic. Cursor has been a trusted

16:33 partner of Anthropic since Sonnet 3.5. What will continue to increase compute to support claude models and cursor and are excited for what comes next with them at Space X? What? This is like the

16:47 most contradictory thing I've ever seen. And I'm going to show you what led to this. And to say that Elon and Anthropic have had a rocky relationship is an understatement. But a few months ago,

17:00 they seem to have patched things up. And it's almost like the enemy of my enemy is my friend kind of situation. So Elon has said pretty horrible things about anthropic. The least of which is just

17:13 calling them misanthropic which means really against humanity. But then they had to work together. Anthropic was in desperate need of additional compute. They had so much demand and they just

17:27 did not have the compute to support that demand. And of course looking over here XAI SpaceX AAI had enormous supply of compute and not enough demand. And this was before the cursor acquisition was

17:41 announced. So here, this is where the relationship gets patched up. Elon says, "I spent a lot of time last week with senior members of the anthropic team to understand what they do to ensure Claude

17:51 is good for humanity. Basically, no one set off my evil detector." And so at that point, all previous friction had been patched up. So that brings us back to this. So forget the rocky

18:03 relationship. forget the fact that Enthropic banned XAI from using its models back in January. Now, no, no, we're all good. You can go ahead and continue to use our models cursor. And

18:14 in fact, this is kind of one of those situations in which Anthropic actually looks kind of good from a PR perspective. Usually, it's open AI coming through when Anthropic makes a

18:24 mistake and kind of picking up the pieces and looking really good like the hero. In this example, it's Anthropic. All right, so now you know all of the history. You know how we got here. You

18:33 see how all of this is so contradictory. Uh but not all is lost in cursor. You can still use your own API key. You can still use the codeex IDE extension and compatible gateways. But what does this

18:48 mean for the future? Well, I'll tell you one thing. Cursor does not want to be dependent on another platform. This is likely why they were acquired. Part of the reason is because they wanted to

19:00 build their own models and XAI had this insane amount of compute that they can use towards building their own models and so now they're building their own models and I think these companies are

19:09 going to become increasingly fearful of being dissolation attacked not only by China but by other US labs and I think increasingly we're not going to see these platforms that support multiple

19:24 model providers that are basically model provider agnostic the cursors the wind surfs of the world. First, the business model is very difficult because you're buying the main thing you need to power

19:35 your platform from somebody else. That is massive platform risk. That's why Windinsurf was acquired. That's probably why again Cursor was acquired. Now Cursor is building their own model. So I

19:45 think what we're going to see and this is really unfortunate is the Open AIs, the XAIs, the Googles, the Anthropics of the world basically going completely verticalized. They're going to build

19:56 their own models. They're going to build their own harnesses. They're going to build their own coding platforms. And if you want to use multiple models in a single place, it's going to become

20:07 increasingly difficult to use multiple models within the same platform. Oh, and if you thought this was going to go down without Elon saying something, here's what he said. I couldn't care less. Scam

20:18 Alman and Greg Stockman, oh man, are utterly untrustworthy who stole an open-source nonprofit. The spite is just so so strong. And one more thing about Tibo from Open AI. I actually did an

20:35 incredible long- form interview with him where he talks about everything from their nextgen models to where agents are going to the reset button. And let me link that right here. Go check it out.