WEBVTT - Monologue: Is Anthropic Killing Cursor? 0:00:02.160 --> 0:00:07.560 Zone Media. Hello and welcome to this week's Better Offline Monologue. 0:00:07.600 --> 0:00:17.520 I'm your host ed zetron over on my premium newsletter. 0:00:17.560 --> 0:00:20.040 This week I reported out a troubling situation with AI 0:00:20.079 --> 0:00:23.239 powered coding environment Cursor maybe, a company called any Sphere 0:00:23.360 --> 0:00:25.800 who unexpectedly changed their pricing in the middle of June 0:00:25.880 --> 0:00:27.639 as well as adding a two hundred dollars a month 0:00:27.760 --> 0:00:31.880 ultra subscription that well pissed everyone off and pissed off 0:00:31.920 --> 0:00:34.240 a lot of people for many different reasons. In fact, 0:00:34.560 --> 0:00:37.879 for some history, Cursor is or was the bell of 0:00:37.880 --> 0:00:40.400 the ball of the AI industry. They had five hundred 0:00:40.400 --> 0:00:43.440 million dollars of annualized recurring revenue, which means their highest 0:00:43.440 --> 0:00:46.880 month times twelve, making them the first high revenue startup 0:00:46.920 --> 0:00:50.080 outside of anthropic and open AI to actually show growth potential. 0:00:50.360 --> 0:00:52.440 There really is no one else like them out there 0:00:52.440 --> 0:00:56.240 at this number. I mean as Alpha sense four hundred 0:00:56.280 --> 0:00:58.840 million according to the AI data based from the information, 0:00:59.320 --> 0:01:01.640 but everyone else is powering around one hundred two hundred 0:01:01.680 --> 0:01:04.920 million arr and raising far more money. Everyone was really 0:01:05.000 --> 0:01:08.120 excited about Cursor. They were growing fast. Everyone loved them, 0:01:08.280 --> 0:01:10.160 and it was because their product allowed you to just 0:01:10.319 --> 0:01:14.720 build software just by typing stuff, and engineers liked it 0:01:14.800 --> 0:01:17.400 and regular people liked it. Not saying it was good 0:01:17.480 --> 0:01:19.640 or anything, just that people liked it, and they liked 0:01:19.680 --> 0:01:22.160 it because you gave you kind of unrestricted access to 0:01:22.160 --> 0:01:25.240 ward these models to build all this stuff. Now, what 0:01:25.280 --> 0:01:28.160 if I told you that perhaps they had grown to 0:01:28.240 --> 0:01:32.280 that five hundred million ARR number using techniques like allowing 0:01:32.319 --> 0:01:34.840 people to spend way more than they were paying for 0:01:34.880 --> 0:01:39.600 the service. Yeah, that's exactly what happened. Previously, Cursor's pricing 0:01:39.640 --> 0:01:42.559 was pretty simple. You had five hundred fast completions, meaning 0:01:42.560 --> 0:01:44.559 that you could ask Cursor to do something using model 0:01:44.600 --> 0:01:47.600 like Claudes sonnet for or Opus four, and it deduct 0:01:47.720 --> 0:01:50.640 requests from your allowance. The amount deductive was based on 0:01:50.680 --> 0:01:52.880 the model, with some models costing less than others, such 0:01:52.920 --> 0:01:56.160 as Anthropics claud for Opus costing more. Once you were 0:01:56.200 --> 0:01:58.240 out of those fast requests, Cursor would then put you 0:01:58.280 --> 0:02:01.000 in the slow lane, the unlimited slow lane, I should add, 0:02:01.120 --> 0:02:03.400 which nevertheless would let you use premium models, but you'd 0:02:03.440 --> 0:02:04.760 be put in the back of the queue and at 0:02:04.760 --> 0:02:07.000 peak times could be waiting some time before you get 0:02:07.040 --> 0:02:09.640 to your completion. You could also select auto and Cursor 0:02:09.680 --> 0:02:12.079 would select whatever was available the best model for the 0:02:12.120 --> 0:02:15.440 job at the time, never want to turn the best 0:02:16.000 --> 0:02:19.280 over to the company though. But on June seventeenth, twenty 0:02:19.320 --> 0:02:21.880 twenty five, Cursor launched that two hundred dollars a month 0:02:22.080 --> 0:02:26.040 ultraplan along with a new and confusing pro package It's 0:02:26.160 --> 0:02:28.880 very muddied. A new wording around said twenty dollars a 0:02:28.880 --> 0:02:31.560 month plan and how it would work going forward, claiming 0:02:31.560 --> 0:02:33.520 that they would be rolling out changes to make it 0:02:33.560 --> 0:02:36.400 more generous, but actually changing pricing to reflect new wording, 0:02:36.400 --> 0:02:39.640 offering unlimited agent requests and no real explanation as to 0:02:39.680 --> 0:02:41.919 what that meant, and no further mention of any request 0:02:41.960 --> 0:02:44.920 limits or what those requests or rate limits might be. 0:02:45.840 --> 0:02:48.760 The reality was, of course, far grimmer. Cursor users have 0:02:48.760 --> 0:02:51.919 found themselves heavily rate limited, especially on Anthropic models, but 0:02:51.960 --> 0:02:55.200 the so called unlimited agent requests mostly pushing towards Cursor's 0:02:55.280 --> 0:02:58.160 on frontier model, which users claim is nowhere near as 0:02:58.200 --> 0:03:02.080 effective as model SMOP and AI Andthropic. On June thirtieth, 0:03:02.440 --> 0:03:05.200 Curser made another change to their product features, changing unlimited 0:03:05.240 --> 0:03:08.360 agents to extended limits on agents, further money in pricing 0:03:08.440 --> 0:03:10.680 on a product that was once renowned for its simplicity 0:03:10.760 --> 0:03:14.919 and lack of limits. So this isn't good, is it. 0:03:15.240 --> 0:03:17.799 Silicon Valley's favorite coding startup has in the last few 0:03:17.800 --> 0:03:20.560 weeks completely changed how its customers interact with its product, 0:03:20.680 --> 0:03:23.040 both degrading the service and making it far more expensive 0:03:23.040 --> 0:03:25.359 in the process. They also allow you now to use 0:03:25.480 --> 0:03:29.359 usage based pricing. Now, this is the funny one. If 0:03:29.360 --> 0:03:31.799 you pay twenty dollars a month on Cursor, they will 0:03:31.840 --> 0:03:35.400 guarantee you twenty dollars a month of AI compute at least, 0:03:35.400 --> 0:03:37.840 so they're literally giving money away. They add a twenty 0:03:37.880 --> 0:03:40.920 percent fee on top of compute. But wow, so you're 0:03:40.920 --> 0:03:43.400 making twenty percent morgin. Now you're fucking not. They're losing 0:03:43.400 --> 0:03:46.800 money on everything. I'm sure of them. But it's strange. 0:03:46.880 --> 0:03:49.560 It's strange, and it's bad. But this is also textbook 0:03:49.560 --> 0:03:52.360 and shitification. Corey doctor O's term for when platforms offer 0:03:52.400 --> 0:03:54.400 a high quality product to gain a large user base, 0:03:54.680 --> 0:03:57.280 usually through convenience or great value. Then the grade the 0:03:57.320 --> 0:03:59.080 service over time to make more money as a means 0:03:59.120 --> 0:04:02.839 of maximizing value shareholders or making money themselves. It's also 0:04:02.920 --> 0:04:05.360 part of my rot economy. Thesis. The growth of all 0:04:05.400 --> 0:04:08.200 costs thinking has dominated the tech industry, and I'd argue, 0:04:08.240 --> 0:04:11.120 thanks to the proliferation of business idi it's controlling everything 0:04:11.720 --> 0:04:14.280 that they've drained some of the logic behind in certification away, 0:04:14.320 --> 0:04:16.320 because very few of these companies actually have a plan 0:04:16.400 --> 0:04:20.400 for sustainability, let alone profitability. Generally, in certification gets people 0:04:20.440 --> 0:04:23.719 through the door and makes the service totally impossible to avoid, 0:04:23.839 --> 0:04:26.960 makes it essential. That's really difficult to do with a 0:04:27.000 --> 0:04:30.839 paid software product like Cursor, because AI compute is so 0:04:31.040 --> 0:04:35.760 fucking expensive. It's so expensive, so Cursor has decided to 0:04:35.839 --> 0:04:39.560 en shittify without making sure that they have something essential. 0:04:40.760 --> 0:04:43.479 But the thing is, I had texted Cory about this. 0:04:43.839 --> 0:04:47.000 I think this actually is something quite different. I think 0:04:47.040 --> 0:04:50.680 this is the world's first chain in shitification. My belief 0:04:50.720 --> 0:04:53.599 is simple. Any Sphere, which is the company that makes Cursor, 0:04:53.600 --> 0:04:56.440 by the way, is despite getting nine hundred million dollars 0:04:56.440 --> 0:04:58.880 in funding in early May, running out of money, or 0:04:58.880 --> 0:05:01.200 at least believes that continue to operate its business in 0:05:01.240 --> 0:05:02.800 the way it did less than a month ago would 0:05:02.800 --> 0:05:05.360 cause it to do so. But why the sudden changes, 0:05:05.400 --> 0:05:07.279 Why the knife in the heart of their customers just 0:05:07.320 --> 0:05:10.599 after raising nearly a billion dollars. It's simple, it's a 0:05:10.680 --> 0:05:13.839 chain and citification. Anthropic jacked up their prices, and so 0:05:13.920 --> 0:05:17.320 did open Ai. On May twenty second, twenty twenty five, 0:05:17.360 --> 0:05:20.000 a few weeks after Cursor raised nine hundred million dollars 0:05:20.000 --> 0:05:23.080 and Thropic launched both claud Sonet and claud Opus four. 0:05:23.480 --> 0:05:26.920 You might say Claude four Opus of Claude force Sonet. 0:05:27.200 --> 0:05:30.640 I don't care. But these were two new powerful as 0:05:30.680 --> 0:05:34.000 judged by benchmarks made specifically for large language models. And 0:05:34.200 --> 0:05:37.800 I should add that the expensive claud Opus four was 0:05:37.839 --> 0:05:41.479 more focused on coding benchmarks. This is an important detail now. 0:05:41.480 --> 0:05:44.200 Eight days later, on May thirtieth, twenty twenty five, patron 0:05:44.240 --> 0:05:48.080 Anthropics API documentation appeared for the first time called service tiers, 0:05:48.440 --> 0:05:51.080 adding priority tiers for enterprise stats that didn't want to 0:05:51.120 --> 0:05:54.440 and I quote sorry, I mean did want to and 0:05:54.520 --> 0:05:57.919 I quote provide a guarantee around the infrequency of server 0:05:58.040 --> 0:06:02.120 overloaded errors even during peak time. Hmm, that's not good. 0:06:02.520 --> 0:06:06.320 Anthropic service tiers require a multi month upfront commitment on 0:06:06.360 --> 0:06:08.400 how many tokens per minute your startup will use, but 0:06:08.480 --> 0:06:12.279 also add an insidious charge around prompt casing. Now Prompt 0:06:12.320 --> 0:06:14.760 cashing is when when you put something into a model, 0:06:14.839 --> 0:06:16.520 such as a code base in the case of a 0:06:16.520 --> 0:06:19.200 coding startup, or I don't know a great deal of 0:06:19.200 --> 0:06:21.880 stuff about how a model might want to act, you 0:06:21.920 --> 0:06:23.479 put it in the cash so that it kind of 0:06:23.520 --> 0:06:25.400 like RAM, it just kind of reads off of it 0:06:25.440 --> 0:06:30.400 instead of rewriting and rewriting and rewriting. So they've added 0:06:30.440 --> 0:06:33.160 a vig to it or attacks whatever you call it, 0:06:33.200 --> 0:06:36.200 a toll perhaps, and they're now charging and this is 0:06:36.279 --> 0:06:38.599 really fucking insidious, either one hundred and twenty five percent 0:06:38.720 --> 0:06:40.800 or two hundred percent of the cost of casing information 0:06:40.880 --> 0:06:43.680 that you need to access it more readily. And I 0:06:43.680 --> 0:06:46.200 should add this is only if you want the Priority tier. 0:06:46.760 --> 0:06:51.080 And another thing, coding startups are extremely prompt cash heavy, 0:06:51.960 --> 0:06:54.800 kind of fucked up right, very fucked up, and they're 0:06:54.839 --> 0:06:57.080 not the only ones doing it. On June twenty fifth, 0:06:57.120 --> 0:07:00.159 twenty twenty five, open Ai also launched Priority processing for 0:07:00.200 --> 0:07:02.760 API customers. They claim it's the payers you go to 0:07:02.839 --> 0:07:05.599 where you pay higher API prices for and I quote 0:07:05.680 --> 0:07:09.800 predictable low latency. However, despite this being payers you go, 0:07:09.880 --> 0:07:12.240 the service is only available for their enterprise customers who 0:07:12.280 --> 0:07:14.960 have made an up prompt commitment much like Anthropic does, 0:07:15.040 --> 0:07:17.080 the difference being they are not trying to tax you 0:07:17.120 --> 0:07:20.800 on prompt caching now. Cursor CEO Michael Truell also said 0:07:20.800 --> 0:07:23.080 when announcing their agree just two hundred dollars a month 0:07:23.160 --> 0:07:26.160 plan that it was only made possible by multi year 0:07:26.160 --> 0:07:30.400 partnerships with OpenAI, Anthropic, Google, and Xai. In short, I 0:07:30.400 --> 0:07:32.920 think that nine hundred million dollars that Cursor got may 0:07:32.960 --> 0:07:35.520 have immediately been handed to or committed to the major 0:07:35.560 --> 0:07:39.600 model developers. They got looted baby. And now what makes 0:07:39.640 --> 0:07:42.760 this weirder is that Cursor is also Anthropic's largest customer, 0:07:42.960 --> 0:07:45.240 with sources claiming that they're such a large customer that 0:07:45.280 --> 0:07:51.800 they're taxing anthropics infrastructure and making them run out of GPUs. Now. 0:07:51.920 --> 0:07:54.920 I should also add that Cursor isn't the only one 0:07:54.960 --> 0:07:58.880 that's recently dramatically changed their prices. Vibe coding startups like 0:07:59.000 --> 0:08:00.680 replt and Lovable Bits both have had to do so 0:08:00.800 --> 0:08:02.840 in the last month. And I think this is just 0:08:02.880 --> 0:08:06.560 the beginning of something really, really, really bad. I believe 0:08:06.560 --> 0:08:09.480 that Anthropic did this either as a deliberate attempt to 0:08:09.520 --> 0:08:12.480 price gouge its largest customers, and or as a means 0:08:12.480 --> 0:08:15.760 of increasing revenue and its money losing software. These changes 0:08:15.760 --> 0:08:18.720 that deliberate aggressive and targeted price increases, and they were 0:08:18.720 --> 0:08:21.080 timed with the launches of claud for Opus and Sonnet, 0:08:21.240 --> 0:08:24.720 which suggests that Anthropics costs have dramatically increased with these models, 0:08:24.720 --> 0:08:27.120 so much that simply increase in the cost per million 0:08:27.120 --> 0:08:30.280 tokens is in sufficient. It's my belief that the launchers 0:08:30.280 --> 0:08:32.520 of both claud Opus and to a lesser extent, son 0:08:32.559 --> 0:08:34.960 It have caused en up hevil in Anthropics costs and 0:08:35.000 --> 0:08:37.480 compute demands, which in turn forced them to start increasing 0:08:37.480 --> 0:08:41.520 costs on their customers. However, Anthropic has I believe, realized 0:08:41.559 --> 0:08:43.480 that there is no real way to just increase the 0:08:43.600 --> 0:08:46.240 cost of Opus and Sonnet for further on just the 0:08:46.240 --> 0:08:48.800 amount of tokens that the customer might use, and that 0:08:48.880 --> 0:08:51.840 doing so might push away smaller developers, which wouldn't make 0:08:51.920 --> 0:08:54.319 them more money. So they decided to find a way 0:08:54.360 --> 0:08:58.400 to specifically exploit the finances of their largest customers coding startups, 0:08:58.440 --> 0:09:00.480 in a way that won't be immediately obvious or that 0:09:00.520 --> 0:09:04.040 would spook non coding assistant customers. Except I'm a crafty 0:09:04.080 --> 0:09:06.520 little fuck and I look at everything all the time. 0:09:06.880 --> 0:09:09.280 I do this for fun. I don't do this because 0:09:09.280 --> 0:09:11.760 I have to listen to me. Wario Amadee. If you're 0:09:11.800 --> 0:09:15.000 listening to this, I'm watching you. I have archive dot 0:09:15.080 --> 0:09:18.360 org and a million diet cokes. Anyway, last year I 0:09:18.400 --> 0:09:21.480 talked about the subprime AI crisis, where almost the entire 0:09:21.480 --> 0:09:23.720 tech industry is bought in on a technology sold at 0:09:23.760 --> 0:09:27.960 this vastly discounted rate, and they've heavily centralized and subsidized 0:09:28.000 --> 0:09:31.320 it too. And I predicted one major thing that these 0:09:31.360 --> 0:09:34.040 model developers would eventually have to find a way to 0:09:34.120 --> 0:09:37.080 make their costs work, that they would have to find 0:09:37.080 --> 0:09:39.760 a way to crank up costs on their customers, because 0:09:39.800 --> 0:09:42.360 otherwise they would just continue burning money. I still think 0:09:42.400 --> 0:09:45.480 they are doing so, and I think that's what's happening here. 0:09:45.840 --> 0:09:48.400 I think anthropic and OpenAI to a lesser extent of 0:09:48.520 --> 0:09:50.480 realized that they need to start making money back on 0:09:50.520 --> 0:09:53.839 these fuckers, and they're doing it. Except the fact that 0:09:53.880 --> 0:09:56.520 the companies of immediate companies like Cursor and Repler have 0:09:56.559 --> 0:09:59.960 immediately had to change their prices suggest that maybe not 0:10:00.000 --> 0:10:03.760 one of these businesses makes sense. Maybe this whole time 0:10:03.920 --> 0:10:07.439 this thing was unsustainable. Maybe it turns out the general 0:10:07.520 --> 0:10:09.960 if AI doesn't have the kind of business returns you 0:10:10.000 --> 0:10:16.079 need to run a startup. If only someone had said something. Remember, 0:10:16.240 --> 0:10:18.520 look see how fucked up saying Remember but I'm not 0:10:18.600 --> 0:10:22.080 editing it out. This is an honest podcast. But seriously, everyone, 0:10:23.080 --> 0:10:28.200 remember the pale horses. Rate limits, service interruptions, price increases, 0:10:28.360 --> 0:10:31.280 trouble raising funding, and trouble with money. The horses are 0:10:31.320 --> 0:10:34.480 drawing nearer and telling you we're coming to the end 0:10:34.480 --> 0:10:36.480 of this. I don't know if it will be soon, 0:10:36.640 --> 0:10:39.040 I don't know if it'll be next year. But nothing 0:10:39.080 --> 0:10:40.920 about this suggests that things are going well.