OpenAI Broke Up With Cursor Transcript, AI Summary & Key Points

00:00 Mom and dad are fighting and I'm going to be real with you guys, it's kind of stressing me out. For those who don't know, I've been a big fan of both Cursor and OpenAI for quite a while. For those who aren't reading the screen right now, what it says is the following. We're ending our partnership with Cursor following its acquisition by SpaceX. Under our proposal, Cursor's direct access to our models would end on November 12th.

00:18 Yeah, this is real. Your Cursor subs will no longer allow you to use OpenAI models. So, instead of spending my weekend relaxing or doing weekend things, I'm stuck here on a Saturday night filming a last second video to cover all of this stuff because it's been, let's be real, kind of chaotic. I've talked to my friends at both Cursor and OpenAI in order to get as many of the important details as I possibly can.

00:39 And I will do my best to make sure you know what's actually going on as well as what you can do to work around these limitations. I have a bad feeling I'm about to piss off both mom and dad. So, let me cover my bases quick with a word from today's sponsor. I have an important question for the builders out there. If you were building an app that was pretty good and useful into a company that was like pretty big came over and asked to use it, you know, like a Fortune 500, maybe a company like a Microsoft or a Google or a

01:03 Stripe, they hit you up and said, "Hey, we want to use your thing. Can you get us set up? Do you actually think you could do it?" Because I have a feeling a lot of you guys are going to get stuck on the first step, which is setting up your app in a way where the company's IT team is willing to onboard. This is where work OS comes in clutch. And there's a reason why everyone from OpenAI and Anthropic to Cursor to Bolt, Clay, Thinking Machines, Fallon, just like look at the logos.

01:26 There's a reason all of these logos are here. And it's not cuz rolling your own off is easy. Rolling your own off is easy. But rolling it in a way that is enterprise ready is not. Especially if you don't want to make your developers hate themselves. Work OS has all of the integrations, SDKs, components, and more that you and your agents expect, which is huge because it means setting it up is not miserable as often can be.

01:45 But more importantly, their admin portal will unblock those big companies from signing up to your app. They set up all of the identity providers for you with this really nice simple panel that you can link your users to. They set it up with their IT teams and they're good to go without you having to do a whole bunch of back and forth to set up their specific SAML Duo ADP or whatever else they're using.

02:03 And if that company's agents need to sign up, they have you covered there, too. Because OMD is a new standard they're helping pioneer with a bunch of other awesome companies. you know, people like Cloudflare, Firecrawl, Resendon, Monday, Kernel, and more in order to try and get a standard that will allow agents to sign up on a user's behalf really ready to go.

02:23 And since they already support it, cuz I mean, obviously they do. They invented it. Works is for everyone, users, enterprises, and agents. Set yourself up for every customer at soy. Oo. I'm going to start by covering exactly what is changing and how it affects things. But I also want to save some time to talk about the personal drama between people involved at the companies.

02:41 The beef forming between Tibo and Michael and the weird conversations that they are dancing around and the details of all that. But most importantly, I want to save some time at the very end to cover how to work around this and make sure you're still getting the best possible models without having to rely on weird business deals like this taking access away from the things you use every day.

03:00 So, let's start with the official post. And there's a little bit at the end I feel like almost everyone has been missing that you should definitely pay attention to. Our decision on Curser following its acquisition by SpaceX. That's an important detail for those who somehow have not caught the news. SpaceX acquired Cursor. It was originally meant to be a joint partnership, but there were two potential outcomes of it.

03:22 Either SpaceX AI would pay Cursor $10 billion for the collab, mostly in order to get access to their data and their engineers around RL to make the Grock models better, which spoiler, the reason Grock got so much better is because they bought Cursor. The other option they had was to at the end of the year acquire Cursor for $60 billion. SpaceX decided to not wait and instead just make the deal happen immediately.

03:42 And the result is Curser is now part of SpaceX. I know this is closing aggressively fast because I probably mentioned this before. I was an early investor in Cursor and I've been getting spammed with emails and still have no idea if I'll even be allowed to get the stock, much less sell it. So, uh, yeah, I might make some money off this deal. I might not.

04:01 It's hard to know. But since the deal is over, so to speak, any allegiance I have to Curser is kind of dead because I do, let's be real, intend to sell that equity as soon as I'm able to from the SpaceX acquisition. But, uh, we'll see what happens when we get there. I have no bottom line tied to how I cover cursor here. Nice to be free of any potential burden I had there, but I'll be honest, I rarely let the fact that I have money in a company prevent me from saying what I want to about them.

04:26 If anything, I tend to be more harsh on the companies that I have financial ties to. I also have no tie financially whatsoever to OpenAI. In fact, I spend a lot of money on OpenAI, so take that as you will. Anthropic is still the company I pay the most. But yeah, anyways, let's read what they said. Today we notified SpaceX that we intend to wind down our contract providing OpenAI models to Cursor with a proposed shut off date of November 12th, 2026.

04:51 To maximize the time that developers can retain access to our models through Curser, we're giving them the maximum notice provided by our contract. This seems to be a internal back and forth between these two companies gone wrong around the details of a specific contract and OpenAI has decided to just honor the exact specification of the contract and not do a renewal or continue extending or whatever other thing that they were doing before.

05:15 The contracts between the companies consuming all of this AI stuff and the companies who are actually making the models and hosting. So the Googles, Anthropics, and OpenAIs of the world versus everyone else tend to be kind of weird because the margins for hosting the models are really good for those businesses, which gives them a lot of wiggle room in how they negotiate these contracts.

05:32 And I think it's important to remember that cursor at a point in time was doing over 70% of the inference on Anthropic. So of all of the stuff Anthropic was doing with their GPUs, 70% of non-ressearch work was going to cursor. So they are a huge spender in the space and as you can guess they have these huge contracts where they negotiate the prices down massively.

05:53 Speaking of cloud models in cursor, the head of compute at anthropic had a post about this too, taking the opportunity to dunk. Curser has been a trusted partner of Anthropic and Sonnet 3.5. We'll continue to increase compute to support cloud models in Cursor and are excited for what comes next with them at SpaceX. You might be surprised here because Anthropic's pretty quick to ban competitors from using their stuff and as I mentioned before, they've already banned XAI before.

06:15 That was before they got a bunch of their compute from SpaceX, though. And now that Anthropic is so reliant on SpaceX for compute, it's not surprising that the compute officer at Anthropic is being nice to them in this beef because Anthropic and XAI might not get along directly, but they both hate OpenAI so much they are more than willing to set aside their differences and partner up here.

06:36 And also the guy who runs computer Anthropic better be thankful for SpaceX because he probably should have been fired before for not acquiring the compute that they needed and giving anthropic the massive disadvantage that they are currently experiencing. It's particularly funny to see the head of comput and anthropic saying that they're working on increasing compute for cursor just a day ago because right afterwards the claude devs accounts announced that they'll be decreasing your limits in cloud code by 17%.

07:00 Don't think what they're saying here is true that it's 25% higher because they're turning off the 50% increase right after. I have a whole video about this coming tomorrow. So, keep an eye out for that. Make sure you're subscribed if you haven't. It cost you nothing. It'll help you keep track of this breaking news as it happens. This is all annoying to keep up with, but I've been doing my best.

07:16 And I bet you that bell can save you a lot of time and money. Anyways, this also means OpenAI is probably losing a pretty big revenue driver by blocking Cursor because Curser doesn't get the absurd subsidization we get in the subscriptions. So you can pay the 200 bucks a month for codeex and get like 8k plus outside of resets in inference for that month.

07:35 Cursor is paying a probably massively discounted rate, but my honest guess is that it would be in the 20 to 30% range, not the 80 to 95% range that we can get with subscriptions. Anyways, back to the details that were shared here. This decision was incredibly tough and we care deeply about our models being broadly available for developers. We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service based on our experience with Elon Musk's company's violating

08:03 contracts. And there's the detail that matters. The thing that OpenAI is concerned about here is that Elon has said on the record that he is distilling OpenAI's models from the information that they get using them. Whether it's internally at their company collecting the inputs and outputs, whether it is the data they can buy from other places or if it is the data that cursor has had for a long time that is clearly quite useful as the models that cursor put out like composer are really good and they got that largely

08:29 from the data that they have from everyone using cursor for all of their AI code for as long as they've been around for. That data advantage has helped cursor a ton with their positioning. And now that they are owned by SpaceX, that data is owned by Elon. And that could be very helpful for Grock. I mean, it is. We've already seen the results. But they want to make sure their competitors aren't getting this huge surge of data from OpenAI.

08:50 For what it is worth, there is a history of this happening in the past. For example, when Windsurf was rumored to be acquired by OpenAI, Anthropic banned access to cloud models in Windsurf concerned that they would use it for distillation. Separately, XAI was banned from using Anthropic models through Cursor. Funny enough, all the way back at the start of this year, again citing concerns that XAI would use the data from anthropic models in cursor in order to RL their models.

09:18 Particularly funny when you realize how close anthropic and XAI are now because Anthropic needed to compute really bad and Elon hates OpenAI so much. But one final one that I think is notable here is that Anthropic revoked OpenAI's access to Claude in their API, claiming that they were violating terms of service when it is generally understood that the use of this API was so that they could do benchmarking of claude models against OpenAI models.

09:43 This is a big part of why anthropic models do not appear as much in OpenAI's benchmarks that they run themselves because they are not allowed to run the anthropic models against benchmarks themselves. They do still include the numbers sometimes, but it's always from external testers because they're not allowed to do the benching themselves. So, there is a precedent for this, but there isn't a precedent for OpenAI doing it, which is really strange.

10:03 It shows just how much they hate Elon and how bad the blood between those companies has come. But, there is one other thing that they are really scared of that we'll get to at the end of this post. We work with a large partner like SpaceX. We typically rely on custom contracts to ensure compliance with our terms of service and that the integration provides for safety at scale.

10:22 After Musk acquired Twitter, now part of SpaceX, the company broke the terms of our contract alongside many others. This is a link to the New York Times, which is particularly funny because of the lawsuit between OpenAI and New York Times. But since I'm not signed in, I can't even get you the title really. Okay, there we go. Nope. I have to refresh over and over to get it.

10:39 Elon Musk ramps up AI efforts even as he warns of danger plans to compete with open AI the chat GBT cool yeah no idea what that is annoying that I can't read it but he more importantly said under oath earlier this year that XAI now also part of SpaceX has violated OpenAI's terms of service terms which are similar to XAI's own again linked here Musk admits to distilling OpenAI data for his XAI business here's why that's controversial you get the idea everyone distills Elon admitted to that I don't think him admitting

11:10 that they're distilling is that big a deal. Data is data and OpenAI and Anthropic had an early advantage by collecting all the data before the web realizes was happening and companies started locking it down. Getting the data from OpenAI and Anthropic by paying them to use their models is in my opinion totally reasonable and it is absurd that either company is trying to ban people from doing explicitly that.

11:30 It is what it is though. I've given anthropic for doing this. I will do the same to OpenAI. It is very bad optics and just generally not good for the ecosystem to be so strict with these things when anthropic and open AI are not as affected because they were so early. In fact, I'd go as far to say the data that SpaceX AI can't get is only inaccessible because of anthropic and open AI.

11:53 So banning them from getting it the other way, which is through the models, is bad. Let's get to the final details of this before I go through the repercussions and how we work around it. Our custom agreement with cursor gives us a limited time window to cancel it after a change of control. As AI capabilities advance, we also have a new level of accountability to ensure for our upcoming model Astra and that it is being used in accordance with our terms.

12:19 Given all of this, we've decided to hold the contract cancellation to the latest date we can while not providing future models to cursor. There's the important detail. This isn't about some new development in OpenAI versus cursor. This is because Astra apparently is a better enough model that OpenAI is concerned about distillation of Astra and as such is making sure any company that isn't compliant with the terms of service is not allowed to get Astra in within their existing contract.

12:48 This does make why OpenAI is doing this more understandable. It doesn't make it better or worse in my opinion. It is just clearer why this is happening because of that. If Astra does turn out to be this massive leap in capability that OpenAI seems to be teasing it will be then our friends at Cursor would benefit greatly from having it in their product and more importantly from getting all of the data.

13:10 Normally you would have to hit the API directly and pay to generate all that. So if I wanted to do I don't know like 100 million tokens I would have to pay for all of that just to get the data to train against. But if I have a customer and they are paying to use the models and then I still get to get the same data, then I'm kind of double dipping because I'm charging a customer to access OpenAI stuff and I get to keep the results so I can use that to train.

13:33 Obviously, if you're clicking the do not use my data button in the cursor dashboard, they're not going to do this. That'd be incredibly illegal. But there's a lot of customers who don't hit that, which ends up being massively useful for Cursor and now SpaceXi in collecting this data to train against. There's a little bit at the bottom here saying that they've worked with Cursor for nearly four years.

13:51 They have enormous respect for the team, the product, yada yada yada. Trying to play nice. A bit over a year ago, I was at the OpenAI office getting early access to GBT5 and filming that video you guys probably saw when it came out. And when they gave us the early access, they gave us a Google doc with detailed instructions on how to set up the model inside of Cursor because that was the best place to test it.

14:09 I would guess that the vast majority of engineers at OpenAI were using Cursor to do their work at the time, even though the Codex CLI already existed. I know for a fact when I was there I didn't see anyone using the codec cli but I did see a lot of people who were using cursor. Thinking back actually I don't know if there were even details on how to use it within the codec cli at the time.

14:27 I vaguely remember it being broken. Regardless cursor has been a core part of both openai and anthropics history and a significant factor in both going so hard on code recently. So, while in many ways this is them shutting out a competitor from access to their models, in others it is a sad end to a long story of OpenAI relying on Cursor and then eventually deciding to compete with them directly and also Elon spending a bunch of money to speed up his own position in the market.

14:54 With all that said, I want to talk about the not great coms I have been seeing from people at OpenAI as well as at Cursor. Here's a post from TBO that I actually think is decent overall. It's kind of crazy to see him posting. If anybody thought that OpenAI's goal here dropping this on a Friday was to suppress the news, I hope that this proves that wrong because it's a reasonable conspiracy if they just posted it on like one small official account and moved on.

15:18 But Tibo did a big quote tweet massively increasing the attention that this post got. We unfortunately have decided that we cannot continue providing access to our models through Cursor and we're ending our partnership. It boils down to trust and we've asked that this takes effect on November 12th to give you some time to plan. Many have used the GBT models through cursor and here are options we know should work in the future.

15:36 First, we will continue to allow using your own openi API key and similarly we will continue to provide access through our IDE extensions for cursor. So if you're using the codeex plugin and cursor I have questions but it'll still work cool separately. We will keep working with the broadest range of tools and harnesses some of which are open source but also many many closed source ones.

15:56 I will say I had a contact at OpenAI hit us up making sure we knew that this will not affect T3 code and that they plan to continue supporting projects like ours indefinitely going forward. So if you're concerned that your codec sub is going to break in T3 code, I have been told you have nothing to worry about. To be fair, I had assumed as much and never even bothered asking, but it was nice to see the productivity.

16:16 So this pose was fine. No issue with how Tibo framed this. I think he did a reasonable job sharing how we can work around this as well as signal boosting this issue rather than trying to suppress it. And I'll be real, of all the people here, Michael's the one I have talked to the most. I've been helping out the Cursor guys for better part of four years now, give or take.

16:36 I remember doing calls with him forever ago, giving them a bunch of about Cursor. And to this day, I'd still consider him a friend. We've done a handful of calls here and there. He has been an awesome supporter of me and all the chaos that we've been doing over on my channel and with my business forever. So, know that this is coming from a place of love and I do genuinely really respect the people at Kurser.

16:55 I think they have one of the best teams in the world. I think this post was unnecessarily dunky, especially because cursor's kind of stuck in the middle here between the beef on Elon and OpenAI. Cursor could have quietly wo me a bit here and been fine, but instead they ramped up the attack, which I don't think was necessary. We're sorry to see that OpenAI put out a note saying that they plan to block cursor users from accessing OpenAI models in three months.

17:20 This is egregious wording in my opinion because you can still use OpenAI models in cursor if you use them through the API or other ways. And if you're a cursor user, there's a good chance you're also an OpenAI or chat GBT user. Obviously, you can still use OpenAI models. A very literal read of this suggests that cursor users are banned from OpenAI, which is not the case at all.

17:40 The next part is even more egregious in my opinion. OpenAI models serve about 5% of Cursor user traffic and we're speaking with the OpenAI team to resolve this. Cursor was one of the very first users of OpenAI. We've worked closely with the team for years and we've trusted their platform to be neutral infrastructure for our business. This is fine. I wish this was the main post, but this sentence in the middle here about the 5% of cursor user traffic, this one's a bit egregious.

18:05 There are lots of layers to why this is egregious. If they had shared more numbers, it would be less so. But the biggest by far is token utilization. There's a lot of different methods you can measure usage for AI. You can measure the number of users who selected a model in a given window. You can measure the number of prompts they send to a model in a given window.

18:21 You can measure the amount of code written by the model. You can do the thing that most do, which is measure the number of tokens. Generally speaking, OpenAI models will always look bad if you measure in terms of the number of tokens going through them compared to other things because they're so much more token efficient. If we look at the benches on artificial analysis, for example, you will see a very clear lead held by anthropic with both opus 5 and fable 5 placing higher than soul.

18:45 While I think Claude Opus 5 is ranked way too high on this, I would honestly say Fable 5 is ranked a little low because the gap between Fable and Soul feels very real in my day-to-day usage and Fable 5 is still my preferred model for actually writing and landing code in real projects. That said, we're not here to talk about the intelligence or even the cost per task, which is pretty absurd when you look at it because the fable tasks are over $3 per run and souls under a dollar per run.

19:10 So, it's a over 3x difference. A big part of that difference comes from the output token efficiency, which is the thing I actually care about here. 5.6 soul is an incredibly effective and efficient model. Even on the max tier, it's pretty efficient, but if you bump down to X high or high, it's even more efficient. And you'll see here that it was about 17k tokens per task to go through the artificial analysis bench, which compared to Fable at 36k tokens per task is absurd.

19:34 That is more than double the tokens per task. If these were just my numbers, that'd be one thing. But thankfully, we have a somewhat neutral source here, cursor themselves. And you can see that Fable is able to score higher, getting as high as a 70.5% in Cursor Bench on Max. And while Soul's only at 67.2%, 2%. It does it with 28k tokens per task versus 103,000 tokens per task.

19:59 A more than 3x gap in real world agentic work. So what's the point of be sharing these token efficiency numbers? You guys have probably heard a lot of this before. Well, it comes down to the tweet from our friend Michael. When he says 5% of cursor user traffic, it is not specified whether or not that 5% is of requests, of users, or more likely of tokens.

20:21 Because if anthropic models are three times less token efficient, then the 5% should be thought of more like 15%. I wish they would give a little bit more info here, but I understand why they're not. They want to just dunk hard. And to their credit, OpenAI did drop this news pretty harshly on a Friday evening, forcing them to scramble to respond. On that note, it's time for me to be a little more harsh to OpenAI because of uh this guy.

20:43 I've enjoyed working with Cursor pre-acquisition and have respect for the team and what they have built. The 5% here should have come with a strong caveat and I would love for Michael to share the math. It's two of us. Tokens are not a proxy for revenue nor value created. And the open 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 they will inflate traffic share significantly.

21:06 Everything Tibo said here is true and it was entirely unnecessary to do such. I don't think he should have made this post. I think it would have been much better for him to not make a personal beef with Michael over at Cursor because that's my job. Let me do this, man. Seriously though, like I don't think it's a great look for people high up at these companies to be the ones doing this beef public because this is affecting users.

21:30 Neither Tibo nor Michael are meaningly affected in their day-to-day work or their livelihoods even by these changes, but they have a massive negative impact on customers. So fighting over the details of how this number was measured, it just kind of feels childish and unnecessary and it just makes the whole thing have a bad vibe that it didn't before.

21:50 Only remaining thing to talk about is what do you do now? If you want to have one subscription for all of the models, I'm going to be real with you guys, you're kind of screwed because the rates you can get from subscribing to the providers directly are just so much better than what you can get subscribing to an all-in-one service. I'll be real with you guys, we've been kind of screwed with this with T3 Chat, and it's a big part of what's motivated us on T3 Code.

22:14 We can't provide the same level of subsidization that the labs do. If their margins are truly in the 70 to 90% range for inference that have been rumored historically, there's no world in which we can compete because they get to serve the same thing for way cheaper than we do. A prompt on T3 chat might cost us $2 and the same prompt on chat GBT costs them 20.

22:32 How do we win there? We can't. And while cursor has been able to float this for a while for a ton of different reasons, now they're being forced to stop. One of the reasons they could justify this was that data. And the data they kept helped them make the Grock models way better. And let's be real, Grock 4.5 and 4.6 are totally usable for a lot of realworld code work.

22:52 They are not Frontier though. So, if you've already gotten a sip of the Kool-Aid and have felt the power possible with models like Fable and Soul, going back to Grock feels like a step back to the stone age in a lot of ways where you have to orchestrate the models yourself, so to speak. You choose how the work is broken up and where it goes and do a lot of careful auditing of that work directly.

23:13 Fable and Soul have changed how I work. And if you are addicted to that, I'm going to be real with you, you probably need to use the subscriptions from the official labs. The value you get for a $200 cloud code sub or $200 codec sub just shits all over what you can get anywhere else. You can get a surprising amount of Grock, like a truly unlimited feeling amount from the plans with cursor and XAI and whatever the current deal is with the Gro heavy whatever.

23:36 And Grockbots's also a very very compelling product that makes the Grock and Cursor ecosystem way more tempting than it might have been in the past. All of that said, it is pretty hard to live without a sub on cloud code or codeex if you want to have frontier performance. Right now, I've done the numbers for my own usage. My handful of subscriptions across both providers cost me like a grand a month and I'm doing upwards of a hundred grand a month in inference.

24:00 So yeah, if you want one plan for everything, I'm sorry, but if you want one app for everything, there's a lot of cool options. I'll be real, I am tempted to just plug T3 code here, but there's lots of other awesome options like Gene from the developer of Coolifi or Superset, an open source terminal focused solution, or even Herder, which is kind of a new flavor of a T-Mux type panel manager that is also open source.

24:24 Yes, believe it or not, all three of these options are open source, which is pretty cool. I wanted to call these out cuz I know a lot of people want to have one UI or platform that they use for all of these things, and I am totally with you there. I was trying some things out in cloud code and Codex today and switching between the two is obnoxious. Having one unified place to do all of this is great.

24:43 Encursor was an awesome option because their harness was good, their app was good enough, their cloud stuff was really good and you got access to every model for a reasonable price, although not as much usage with them. If you were willing to fall back on Grock, it was pretty solid. Now you're just kind of screwed and have to maintain subs with the different platforms if you want to get the best usage out of them all.

25:03 So yeah, check out all these options and of course know that T3 Code is available whenever you want it. It's free, open source, works great with remote, yada yada yada. We're open source, so it's very unlikely that we're going to get banned compared to other things. So yeah, if you want to avoid getting hurt by companies beefing with each other, make sure you own your tools and that your relations with the products that you rely on are direct and not routed through other layers like this.

25:26 I have a feeling this is not a one-off thing. If anything, it's the start of the end. We're going to see more and more moves like this from Anthropic, OpenAI, and probably even companies like Google and SpaceX. Speaking of which, hey Google, what the I'm thankful your models are the worst because you're the only company that won't let us integrate your subscriptions into T3 Code without risking our users getting their Google accounts banned.

25:46 You have the accounts that are the most important. Like, if I lost my Google account, that would be lifeending for me. But at the same time, you are the platform that is the most aggressive with banning when people use their subscription in other places because you have the worst possible harness, the worst graphics app for it. Like, like seriously, how do you let Claude Code beat you with a desktop app that barely functions?

26:03 Because anti-gravity is sinfully bad. I ah, sorry, I've been holding that one in for a while. I'm pissed at Google. And the fact that people want anti-gravity in T3 code and we can't do it because Google's models are not the thing furthest behind at the company is hilarious. Okay, sorry. Just had to get that out of me. I've said all I have to on this one.

26:23 This sucks, especially for somebody who likes both of these companies and the models provided by both of these companies. It'd be nice if Cursor could be a one-stop shop as it always has been. And it sucks that OpenAI will not let that happen because of Elon's beef with them. It is what it is. Astra's probably going to be great, and I'm excited for the future.

26:38 But god damn, this moment sucks. And for everybody relying on Cursor for access to all the models. And to those of y'all who I know use 56 Soul in Cursor every day, I'm sorry. This just sucks. There is no ifs, ends, or buts about it. You will not get access to Astra in Cursor, and you will lose access to all OpenAI models in November as well. This sucks. It is what it is. I'm sorry. I hope this update was useful and until next time, peace nerds.